cbor2

Summary

• The cbor2 library is vulnerable to a Denial of Service (DoS) attack caused by uncontrolled recursion when decoding deeply nested CBOR structures.
• This vulnerability affects both the pure Python implementation and the C extension (_cbor2). The C extension correctly uses Python's C-API for recursion protection (Py_EnterRecursiveCall), but this mechanism is designed to prevent a stack overflow by raising a RecursionError. In some environments, this exception is not caught, thus causing the service process to terminate.
• While the library handles moderate nesting, it lacks a configurable, data-driven depth limit independent of Python's global recursion setting. An attacker can supply a crafted CBOR payload containing thousands of nested arrays (e.g., 0x81). When cbor2.loads() attempts to parse this, it hits the interpreter's recursion limit, causing the call to raise a RecursionError.
• By sending a stream of small (<100KB) malicious packets, an attacker can repeatedly crash worker processes faster than they can be restarted, resulting in a complete and sustained Denial of Service.

Details

• The vulnerability stems from the recursive design of the CBORDecoder class, specifically how it decodes nested container types like Arrays and Maps.
• Inside decode_array (and similarly decode_map), the decoder iterates through the number of elements specified in the CBOR header. For each element, it calls self.decode() again to parse the nested item. This recursive call lacks a depth-tracking mechanism.
• Vulnerable Code Locations:
  • cbor2/decoder.py (Pure Python implementation)
  • source/decoder.c (C extension implementation)
• Execution Flow:
  1. The cbor2.loads() function initializes a CBORDecoder and calls its decode() method.
  2. The decode() method reads the initial byte and dispatches control to a specific handler based on the major type. For an Array (Major Type 4), it calls decode_array.
  3. decode_array loops and calls self.decode() for each item, leading to deep recursion when parsing a payload like [...[...[1]...]...].

PoC

import cbor2

DEPTH = 1000

payload = b'\x81' * DEPTH + b'\x01'
print(f"[*] Payload size: {len(payload) / 1024:.2f} KB")
print("[*] Triggering decoder...")

try:
    cbor2.loads(payload)
    print("[+] Parsed successfully (Not Vulnerable)")
except RecursionError:
    print("\n[!] VULNERABLE: RecursionError triggered!")
except Exception as e:
    print(f"\n[-] Unexpected Error: {type(e).__name__}: {e}")

Impact

• Scope: This vulnerability affects any application using cbor2 to parse untrusted data. Common use cases include IoT data processing, WebAuthn (FIDO2) authentication flows, and inter-service communication over COSE (CBOR Object Signing and Encryption).
• Attack Vector: A remote, unauthenticated attacker can achieve a full Denial of Service with a highly efficient, low-bandwidth attack. A payload under 100KB is sufficient to reliably terminate a Python worker process.

Credit

This issue was discovered by Kevin Tu of TMIR at ByteDance. The patch was developed by @agronholm.

Summary

• The cbor2 library is vulnerable to a Denial of Service (DoS) attack caused by uncontrolled recursion when decoding deeply nested CBOR structures.
• This vulnerability affects both the pure Python implementation and the C extension (_cbor2). The C extension correctly uses Python's C-API for recursion protection (Py_EnterRecursiveCall), but this mechanism is designed to prevent a stack overflow by raising a RecursionError. In some environments, this exception is not caught, thus causing the service process to terminate.
• While the library handles moderate nesting, it lacks a configurable, data-driven depth limit independent of Python's global recursion setting. An attacker can supply a crafted CBOR payload containing thousands of nested arrays (e.g., 0x81). When cbor2.loads() attempts to parse this, it hits the interpreter's recursion limit, causing the call to raise a RecursionError.
• By sending a stream of small (<100KB) malicious packets, an attacker can repeatedly crash worker processes faster than they can be restarted, resulting in a complete and sustained Denial of Service.

Details

• The vulnerability stems from the recursive design of the CBORDecoder class, specifically how it decodes nested container types like Arrays and Maps.
• Inside decode_array (and similarly decode_map), the decoder iterates through the number of elements specified in the CBOR header. For each element, it calls self.decode() again to parse the nested item. This recursive call lacks a depth-tracking mechanism.
• Vulnerable Code Locations:
  • cbor2/decoder.py (Pure Python implementation)
  • source/decoder.c (C extension implementation)
• Execution Flow:
  1. The cbor2.loads() function initializes a CBORDecoder and calls its decode() method.
  2. The decode() method reads the initial byte and dispatches control to a specific handler based on the major type. For an Array (Major Type 4), it calls decode_array.
  3. decode_array loops and calls self.decode() for each item, leading to deep recursion when parsing a payload like [...[...[1]...]...].

PoC

import cbor2

DEPTH = 1000

payload = b'\x81' * DEPTH + b'\x01'
print(f"[*] Payload size: {len(payload) / 1024:.2f} KB")
print("[*] Triggering decoder...")

try:
    cbor2.loads(payload)
    print("[+] Parsed successfully (Not Vulnerable)")
except RecursionError:
    print("\n[!] VULNERABLE: RecursionError triggered!")
except Exception as e:
    print(f"\n[-] Unexpected Error: {type(e).__name__}: {e}")

Impact

• Scope: This vulnerability affects any application using cbor2 to parse untrusted data. Common use cases include IoT data processing, WebAuthn (FIDO2) authentication flows, and inter-service communication over COSE (CBOR Object Signing and Encryption).
• Attack Vector: A remote, unauthenticated attacker can achieve a full Denial of Service with a highly efficient, low-bandwidth attack. A payload under 100KB is sufficient to reliably terminate a Python worker process.

Credit

This issue was discovered by Kevin Tu of TMIR at ByteDance. The patch was developed by @agronholm.

Summary

When a CBORDecoder instance is reused across multiple decode operations, values marked with the shareable tag (28) persist in memory and can be accessed by subsequent CBOR messages using the sharedref tag (29). This allows an attacker-controlled message to read data from previously decoded messages if the decoder is reused across trust boundaries.

Details

The issue is in the decoder's handling of the shareables list, which stores values tagged with CBOR tag 28 (shareable) for later reference by tag 29 (sharedref).

When decode_from_bytes() is called or when .fp is set to a new stream, the shareables list is not cleared. This allows references to persist across separate decode operations.

The issue exists in both the C extension and the pure Python decoder.

In the C extension (source/decoder.c), the _CBORDecoder_set_fp function (line ~202) updates the file pointer but does not reset the shareables state:

  static int
  _CBORDecoder_set_fp(CBORDecoderObject *self, PyObject *value, void *closure)
  {
      // ... validation ...
      tmp = self->read;
      self->read = read;
      Py_DECREF(tmp);
      return 0;
      // Missing: PyList_Clear(self->shareables) or equivalent
  }

In the pure Python decoder (cbor2/_decoder.py), the fp setter similarly fails to clear self._shareables.

Similarly, decode_from_bytes() in both implementations saves and restores the read pointer but does not clear the shareables list between decodes.

The shareable/sharedref tags are defined in the CBOR value sharing extension (http://cbor.schmorp.de/value-sharing) with scope limited to a single CBOR data item, not across separate messages.

PoC

import cbor2
from io import BytesIO

# Message from trusted source containing a shareable value
msg1 = cbor2.dumps(cbor2.CBORTag(28, "secret"))

# Attacker-controlled message referencing index 0
msg2 = cbor2.dumps(cbor2.CBORTag(29, 0))

# Decoder reused across trust boundaries
decoder = cbor2.CBORDecoder(BytesIO(b''))
decoder.decode_from_bytes(msg1)
print(decoder.decode_from_bytes(msg2))  # prints "secret"

No special configuration required. Affects any application that reuses a CBORDecoder instance to decode messages from different sources.

Impact

Information disclosure. Applications that reuse a CBORDecoder across trust boundaries are vulnerable if the trusted messages use value sharing (tag 28) and an attacker can send messages containing shared references (tag 29). An attacker who can send a crafted CBOR message containing a sharedref tag can read values from previously decoded messages, potentially exposing sensitive data such as credentials, tokens, or private user data.

Related

A similar issue in the encoder could produce invalid CBOR with dangling shared references:

import cbor2
from io import BytesIO

# Create encoder with value sharing enabled
encoder = cbor2.CBOREncoder(BytesIO(), value_sharing=True)

# Persistent object that will be encoded multiple times
shared_obj = ['hello']

# First encode: array containing shared_obj twice
encoder.encode([shared_obj, shared_obj])
print(f'First encode: {encoder.fp.getvalue().hex()}')
# Output: d81c82d81c816568656c6c6fd81d01

# Second encode: just shared_obj
encoder.fp = BytesIO()
encoder.encode(shared_obj)
result = encoder.fp.getvalue()
print(f'Second encode: {result.hex()}')
# Output: d81d01  (just a shared reference to index 1!)

# Try to decode the second result as standalone CBOR
decoder = cbor2.CBORDecoder(BytesIO(result))
decoded = decoder.decode()
# FAILS: shared reference 1 not found

While primarily a correctness bug, it could cause denial of service if invalid CBOR is transmitted to downstream systems that fail to parse it, or cause silent data corruption if the dangling reference happens to resolve to an unrelated value.

It can also be considered a memory leak in both the decoder and encoder as references are held that will never be released as long as the decoder/encoder remains alive.

Suggested resolution

Add dedicated boolean flags to track when an encode/decode operation is in progress. Reset shared state only when the flag is False (top-level call). This ensures state is reset for standalone calls while preserving shared references for nested calls from hooks (which need access to the registry for cyclic structures).

Decoder (_decoding flag):

• decode(): set flag True, reset state, decode, set flag False
• decode_from_bytes(): reset state only when flag is False

Encoder (_encoding flag):

• encode(): set flag True, reset state, encode, set flag False
• encode_to_bytes(): reset state only when flag is False

Summary

• The cbor2 library is vulnerable to a Denial of Service (DoS) attack caused by uncontrolled recursion when decoding deeply nested CBOR structures.
• This vulnerability affects both the pure Python implementation and the C extension (_cbor2). The C extension correctly uses Python's C-API for recursion protection (Py_EnterRecursiveCall), but this mechanism is designed to prevent a stack overflow by raising a RecursionError. In some environments, this exception is not caught, thus causing the service process to terminate.
• While the library handles moderate nesting, it lacks a configurable, data-driven depth limit independent of Python's global recursion setting. An attacker can supply a crafted CBOR payload containing thousands of nested arrays (e.g., 0x81). When cbor2.loads() attempts to parse this, it hits the interpreter's recursion limit, causing the call to raise a RecursionError.
• By sending a stream of small (<100KB) malicious packets, an attacker can repeatedly crash worker processes faster than they can be restarted, resulting in a complete and sustained Denial of Service.

Details

• The vulnerability stems from the recursive design of the CBORDecoder class, specifically how it decodes nested container types like Arrays and Maps.
• Inside decode_array (and similarly decode_map), the decoder iterates through the number of elements specified in the CBOR header. For each element, it calls self.decode() again to parse the nested item. This recursive call lacks a depth-tracking mechanism.
• Vulnerable Code Locations:
  • cbor2/decoder.py (Pure Python implementation)
  • source/decoder.c (C extension implementation)
• Execution Flow:
  1. The cbor2.loads() function initializes a CBORDecoder and calls its decode() method.
  2. The decode() method reads the initial byte and dispatches control to a specific handler based on the major type. For an Array (Major Type 4), it calls decode_array.
  3. decode_array loops and calls self.decode() for each item, leading to deep recursion when parsing a payload like [...[...[1]...]...].

PoC

import cbor2

DEPTH = 1000

payload = b'\x81' * DEPTH + b'\x01'
print(f"[*] Payload size: {len(payload) / 1024:.2f} KB")
print("[*] Triggering decoder...")

try:
    cbor2.loads(payload)
    print("[+] Parsed successfully (Not Vulnerable)")
except RecursionError:
    print("\n[!] VULNERABLE: RecursionError triggered!")
except Exception as e:
    print(f"\n[-] Unexpected Error: {type(e).__name__}: {e}")

Impact

• Scope: This vulnerability affects any application using cbor2 to parse untrusted data. Common use cases include IoT data processing, WebAuthn (FIDO2) authentication flows, and inter-service communication over COSE (CBOR Object Signing and Encryption).
• Attack Vector: A remote, unauthenticated attacker can achieve a full Denial of Service with a highly efficient, low-bandwidth attack. A payload under 100KB is sufficient to reliably terminate a Python worker process.

Credit

This issue was discovered by Kevin Tu of TMIR at ByteDance. The patch was developed by @agronholm.

Summary

When a CBORDecoder instance is reused across multiple decode operations, values marked with the shareable tag (28) persist in memory and can be accessed by subsequent CBOR messages using the sharedref tag (29). This allows an attacker-controlled message to read data from previously decoded messages if the decoder is reused across trust boundaries.

Details

The issue is in the decoder's handling of the shareables list, which stores values tagged with CBOR tag 28 (shareable) for later reference by tag 29 (sharedref).

When decode_from_bytes() is called or when .fp is set to a new stream, the shareables list is not cleared. This allows references to persist across separate decode operations.

The issue exists in both the C extension and the pure Python decoder.

In the C extension (source/decoder.c), the _CBORDecoder_set_fp function (line ~202) updates the file pointer but does not reset the shareables state:

  static int
  _CBORDecoder_set_fp(CBORDecoderObject *self, PyObject *value, void *closure)
  {
      // ... validation ...
      tmp = self->read;
      self->read = read;
      Py_DECREF(tmp);
      return 0;
      // Missing: PyList_Clear(self->shareables) or equivalent
  }

In the pure Python decoder (cbor2/_decoder.py), the fp setter similarly fails to clear self._shareables.

Similarly, decode_from_bytes() in both implementations saves and restores the read pointer but does not clear the shareables list between decodes.

The shareable/sharedref tags are defined in the CBOR value sharing extension (http://cbor.schmorp.de/value-sharing) with scope limited to a single CBOR data item, not across separate messages.

PoC

import cbor2
from io import BytesIO

# Message from trusted source containing a shareable value
msg1 = cbor2.dumps(cbor2.CBORTag(28, "secret"))

# Attacker-controlled message referencing index 0
msg2 = cbor2.dumps(cbor2.CBORTag(29, 0))

# Decoder reused across trust boundaries
decoder = cbor2.CBORDecoder(BytesIO(b''))
decoder.decode_from_bytes(msg1)
print(decoder.decode_from_bytes(msg2))  # prints "secret"

No special configuration required. Affects any application that reuses a CBORDecoder instance to decode messages from different sources.

Impact

Information disclosure. Applications that reuse a CBORDecoder across trust boundaries are vulnerable if the trusted messages use value sharing (tag 28) and an attacker can send messages containing shared references (tag 29). An attacker who can send a crafted CBOR message containing a sharedref tag can read values from previously decoded messages, potentially exposing sensitive data such as credentials, tokens, or private user data.

Related

A similar issue in the encoder could produce invalid CBOR with dangling shared references:

import cbor2
from io import BytesIO

# Create encoder with value sharing enabled
encoder = cbor2.CBOREncoder(BytesIO(), value_sharing=True)

# Persistent object that will be encoded multiple times
shared_obj = ['hello']

# First encode: array containing shared_obj twice
encoder.encode([shared_obj, shared_obj])
print(f'First encode: {encoder.fp.getvalue().hex()}')
# Output: d81c82d81c816568656c6c6fd81d01

# Second encode: just shared_obj
encoder.fp = BytesIO()
encoder.encode(shared_obj)
result = encoder.fp.getvalue()
print(f'Second encode: {result.hex()}')
# Output: d81d01  (just a shared reference to index 1!)

# Try to decode the second result as standalone CBOR
decoder = cbor2.CBORDecoder(BytesIO(result))
decoded = decoder.decode()
# FAILS: shared reference 1 not found

While primarily a correctness bug, it could cause denial of service if invalid CBOR is transmitted to downstream systems that fail to parse it, or cause silent data corruption if the dangling reference happens to resolve to an unrelated value.

It can also be considered a memory leak in both the decoder and encoder as references are held that will never be released as long as the decoder/encoder remains alive.

Suggested resolution

Add dedicated boolean flags to track when an encode/decode operation is in progress. Reset shared state only when the flag is False (top-level call). This ensures state is reset for standalone calls while preserving shared references for nested calls from hooks (which need access to the registry for cyclic structures).

Decoder (_decoding flag):

• decode(): set flag True, reset state, decode, set flag False
• decode_from_bytes(): reset state only when flag is False

Encoder (_encoding flag):

• encode(): set flag True, reset state, encode, set flag False
• encode_to_bytes(): reset state only when flag is False

Multiple vulnerabilities exist in cbor2 through version 5.7.0 in the decode_definite_long_string() function of the C extension decoder (source/decoder.c):

1. Integer Underflow Leading to Out-of-Bounds Read (CWE-191, CWE-125): An incorrect variable reference and missing state reset in the chunk processing loop causes buffer_length to not be reset to zero after UTF-8 character consumption. This results in subsequent chunk_length calculations producing negative values (e.g., chunk_length = 65536 - buffer_length), which are passed as signed integers to the read() method, potentially triggering unlimited read operations and resource exhaustion.
2. Memory Leak via Missing Reference Count Release (CWE-401): The main processing loop fails to release Python object references (Py_DECREF) for chunk objects allocated in each iteration. For CBOR strings longer than 65536 bytes, this causes cumulative memory leaks proportional to the payload size, enabling memory exhaustion attacks through repeated processing of large CBOR payloads.

Both vulnerabilities can be exploited remotely without authentication by sending specially-crafted CBOR data containing definite-length text strings with multi-byte UTF-8 characters positioned at 65536-byte chunk boundaries. Successful exploitation results in denial of service through process crashes (CBORDecodeEOF exceptions) or memory exhaustion. The vulnerabilities affect all applications using cbor2's C extension to process untrusted CBOR data, including web APIs, IoT data collectors, and message queue processors. Fixed in commit 851473490281f82d82560b2368284ef33cf6e8f9 pushed with released version 5.7.1.

Summary

• The cbor2 library is vulnerable to a Denial of Service (DoS) attack caused by uncontrolled recursion when decoding deeply nested CBOR structures.
• This vulnerability affects both the pure Python implementation and the C extension (_cbor2). The C extension correctly uses Python's C-API for recursion protection (Py_EnterRecursiveCall), but this mechanism is designed to prevent a stack overflow by raising a RecursionError. In some environments, this exception is not caught, thus causing the service process to terminate.
• While the library handles moderate nesting, it lacks a configurable, data-driven depth limit independent of Python's global recursion setting. An attacker can supply a crafted CBOR payload containing thousands of nested arrays (e.g., 0x81). When cbor2.loads() attempts to parse this, it hits the interpreter's recursion limit, causing the call to raise a RecursionError.
• By sending a stream of small (<100KB) malicious packets, an attacker can repeatedly crash worker processes faster than they can be restarted, resulting in a complete and sustained Denial of Service.

Details

• The vulnerability stems from the recursive design of the CBORDecoder class, specifically how it decodes nested container types like Arrays and Maps.
• Inside decode_array (and similarly decode_map), the decoder iterates through the number of elements specified in the CBOR header. For each element, it calls self.decode() again to parse the nested item. This recursive call lacks a depth-tracking mechanism.
• Vulnerable Code Locations:
  • cbor2/decoder.py (Pure Python implementation)
  • source/decoder.c (C extension implementation)
• Execution Flow:
  1. The cbor2.loads() function initializes a CBORDecoder and calls its decode() method.
  2. The decode() method reads the initial byte and dispatches control to a specific handler based on the major type. For an Array (Major Type 4), it calls decode_array.
  3. decode_array loops and calls self.decode() for each item, leading to deep recursion when parsing a payload like [...[...[1]...]...].

PoC

import cbor2

DEPTH = 1000

payload = b'\x81' * DEPTH + b'\x01'
print(f"[*] Payload size: {len(payload) / 1024:.2f} KB")
print("[*] Triggering decoder...")

try:
    cbor2.loads(payload)
    print("[+] Parsed successfully (Not Vulnerable)")
except RecursionError:
    print("\n[!] VULNERABLE: RecursionError triggered!")
except Exception as e:
    print(f"\n[-] Unexpected Error: {type(e).__name__}: {e}")

Impact

• Scope: This vulnerability affects any application using cbor2 to parse untrusted data. Common use cases include IoT data processing, WebAuthn (FIDO2) authentication flows, and inter-service communication over COSE (CBOR Object Signing and Encryption).
• Attack Vector: A remote, unauthenticated attacker can achieve a full Denial of Service with a highly efficient, low-bandwidth attack. A payload under 100KB is sufficient to reliably terminate a Python worker process.

Credit

This issue was discovered by Kevin Tu of TMIR at ByteDance. The patch was developed by @agronholm.

Summary

When a CBORDecoder instance is reused across multiple decode operations, values marked with the shareable tag (28) persist in memory and can be accessed by subsequent CBOR messages using the sharedref tag (29). This allows an attacker-controlled message to read data from previously decoded messages if the decoder is reused across trust boundaries.

Details

The issue is in the decoder's handling of the shareables list, which stores values tagged with CBOR tag 28 (shareable) for later reference by tag 29 (sharedref).

When decode_from_bytes() is called or when .fp is set to a new stream, the shareables list is not cleared. This allows references to persist across separate decode operations.

The issue exists in both the C extension and the pure Python decoder.

In the C extension (source/decoder.c), the _CBORDecoder_set_fp function (line ~202) updates the file pointer but does not reset the shareables state:

  static int
  _CBORDecoder_set_fp(CBORDecoderObject *self, PyObject *value, void *closure)
  {
      // ... validation ...
      tmp = self->read;
      self->read = read;
      Py_DECREF(tmp);
      return 0;
      // Missing: PyList_Clear(self->shareables) or equivalent
  }

In the pure Python decoder (cbor2/_decoder.py), the fp setter similarly fails to clear self._shareables.

Similarly, decode_from_bytes() in both implementations saves and restores the read pointer but does not clear the shareables list between decodes.

The shareable/sharedref tags are defined in the CBOR value sharing extension (http://cbor.schmorp.de/value-sharing) with scope limited to a single CBOR data item, not across separate messages.

PoC

import cbor2
from io import BytesIO

# Message from trusted source containing a shareable value
msg1 = cbor2.dumps(cbor2.CBORTag(28, "secret"))

# Attacker-controlled message referencing index 0
msg2 = cbor2.dumps(cbor2.CBORTag(29, 0))

# Decoder reused across trust boundaries
decoder = cbor2.CBORDecoder(BytesIO(b''))
decoder.decode_from_bytes(msg1)
print(decoder.decode_from_bytes(msg2))  # prints "secret"

No special configuration required. Affects any application that reuses a CBORDecoder instance to decode messages from different sources.

Impact

Information disclosure. Applications that reuse a CBORDecoder across trust boundaries are vulnerable if the trusted messages use value sharing (tag 28) and an attacker can send messages containing shared references (tag 29). An attacker who can send a crafted CBOR message containing a sharedref tag can read values from previously decoded messages, potentially exposing sensitive data such as credentials, tokens, or private user data.

Related

A similar issue in the encoder could produce invalid CBOR with dangling shared references:

import cbor2
from io import BytesIO

# Create encoder with value sharing enabled
encoder = cbor2.CBOREncoder(BytesIO(), value_sharing=True)

# Persistent object that will be encoded multiple times
shared_obj = ['hello']

# First encode: array containing shared_obj twice
encoder.encode([shared_obj, shared_obj])
print(f'First encode: {encoder.fp.getvalue().hex()}')
# Output: d81c82d81c816568656c6c6fd81d01

# Second encode: just shared_obj
encoder.fp = BytesIO()
encoder.encode(shared_obj)
result = encoder.fp.getvalue()
print(f'Second encode: {result.hex()}')
# Output: d81d01  (just a shared reference to index 1!)

# Try to decode the second result as standalone CBOR
decoder = cbor2.CBORDecoder(BytesIO(result))
decoded = decoder.decode()
# FAILS: shared reference 1 not found

While primarily a correctness bug, it could cause denial of service if invalid CBOR is transmitted to downstream systems that fail to parse it, or cause silent data corruption if the dangling reference happens to resolve to an unrelated value.

It can also be considered a memory leak in both the decoder and encoder as references are held that will never be released as long as the decoder/encoder remains alive.

Suggested resolution

Add dedicated boolean flags to track when an encode/decode operation is in progress. Reset shared state only when the flag is False (top-level call). This ensures state is reset for standalone calls while preserving shared references for nested calls from hooks (which need access to the registry for cyclic structures).

Decoder (_decoding flag):

• decode(): set flag True, reset state, decode, set flag False
• decode_from_bytes(): reset state only when flag is False

Encoder (_encoding flag):

• encode(): set flag True, reset state, encode, set flag False
• encode_to_bytes(): reset state only when flag is False

Multiple vulnerabilities exist in cbor2 through version 5.7.0 in the decode_definite_long_string() function of the C extension decoder (source/decoder.c):

1. Integer Underflow Leading to Out-of-Bounds Read (CWE-191, CWE-125): An incorrect variable reference and missing state reset in the chunk processing loop causes buffer_length to not be reset to zero after UTF-8 character consumption. This results in subsequent chunk_length calculations producing negative values (e.g., chunk_length = 65536 - buffer_length), which are passed as signed integers to the read() method, potentially triggering unlimited read operations and resource exhaustion.
2. Memory Leak via Missing Reference Count Release (CWE-401): The main processing loop fails to release Python object references (Py_DECREF) for chunk objects allocated in each iteration. For CBOR strings longer than 65536 bytes, this causes cumulative memory leaks proportional to the payload size, enabling memory exhaustion attacks through repeated processing of large CBOR payloads.

Both vulnerabilities can be exploited remotely without authentication by sending specially-crafted CBOR data containing definite-length text strings with multi-byte UTF-8 characters positioned at 65536-byte chunk boundaries. Successful exploitation results in denial of service through process crashes (CBORDecodeEOF exceptions) or memory exhaustion. The vulnerabilities affect all applications using cbor2's C extension to process untrusted CBOR data, including web APIs, IoT data collectors, and message queue processors. Fixed in commit 851473490281f82d82560b2368284ef33cf6e8f9 pushed with released version 5.7.1.

Summary

• The cbor2 library is vulnerable to a Denial of Service (DoS) attack caused by uncontrolled recursion when decoding deeply nested CBOR structures.
• This vulnerability affects both the pure Python implementation and the C extension (_cbor2). The C extension correctly uses Python's C-API for recursion protection (Py_EnterRecursiveCall), but this mechanism is designed to prevent a stack overflow by raising a RecursionError. In some environments, this exception is not caught, thus causing the service process to terminate.
• While the library handles moderate nesting, it lacks a configurable, data-driven depth limit independent of Python's global recursion setting. An attacker can supply a crafted CBOR payload containing thousands of nested arrays (e.g., 0x81). When cbor2.loads() attempts to parse this, it hits the interpreter's recursion limit, causing the call to raise a RecursionError.
• By sending a stream of small (<100KB) malicious packets, an attacker can repeatedly crash worker processes faster than they can be restarted, resulting in a complete and sustained Denial of Service.

Details

• The vulnerability stems from the recursive design of the CBORDecoder class, specifically how it decodes nested container types like Arrays and Maps.
• Inside decode_array (and similarly decode_map), the decoder iterates through the number of elements specified in the CBOR header. For each element, it calls self.decode() again to parse the nested item. This recursive call lacks a depth-tracking mechanism.
• Vulnerable Code Locations:
  • cbor2/decoder.py (Pure Python implementation)
  • source/decoder.c (C extension implementation)
• Execution Flow:
  1. The cbor2.loads() function initializes a CBORDecoder and calls its decode() method.
  2. The decode() method reads the initial byte and dispatches control to a specific handler based on the major type. For an Array (Major Type 4), it calls decode_array.
  3. decode_array loops and calls self.decode() for each item, leading to deep recursion when parsing a payload like [...[...[1]...]...].

PoC

import cbor2

DEPTH = 1000

payload = b'\x81' * DEPTH + b'\x01'
print(f"[*] Payload size: {len(payload) / 1024:.2f} KB")
print("[*] Triggering decoder...")

try:
    cbor2.loads(payload)
    print("[+] Parsed successfully (Not Vulnerable)")
except RecursionError:
    print("\n[!] VULNERABLE: RecursionError triggered!")
except Exception as e:
    print(f"\n[-] Unexpected Error: {type(e).__name__}: {e}")

Impact

• Scope: This vulnerability affects any application using cbor2 to parse untrusted data. Common use cases include IoT data processing, WebAuthn (FIDO2) authentication flows, and inter-service communication over COSE (CBOR Object Signing and Encryption).
• Attack Vector: A remote, unauthenticated attacker can achieve a full Denial of Service with a highly efficient, low-bandwidth attack. A payload under 100KB is sufficient to reliably terminate a Python worker process.

Credit

This issue was discovered by Kevin Tu of TMIR at ByteDance. The patch was developed by @agronholm.

Summary

When a CBORDecoder instance is reused across multiple decode operations, values marked with the shareable tag (28) persist in memory and can be accessed by subsequent CBOR messages using the sharedref tag (29). This allows an attacker-controlled message to read data from previously decoded messages if the decoder is reused across trust boundaries.

Details

The issue is in the decoder's handling of the shareables list, which stores values tagged with CBOR tag 28 (shareable) for later reference by tag 29 (sharedref).

When decode_from_bytes() is called or when .fp is set to a new stream, the shareables list is not cleared. This allows references to persist across separate decode operations.

The issue exists in both the C extension and the pure Python decoder.

In the C extension (source/decoder.c), the _CBORDecoder_set_fp function (line ~202) updates the file pointer but does not reset the shareables state:

  static int
  _CBORDecoder_set_fp(CBORDecoderObject *self, PyObject *value, void *closure)
  {
      // ... validation ...
      tmp = self->read;
      self->read = read;
      Py_DECREF(tmp);
      return 0;
      // Missing: PyList_Clear(self->shareables) or equivalent
  }

In the pure Python decoder (cbor2/_decoder.py), the fp setter similarly fails to clear self._shareables.

Similarly, decode_from_bytes() in both implementations saves and restores the read pointer but does not clear the shareables list between decodes.

The shareable/sharedref tags are defined in the CBOR value sharing extension (http://cbor.schmorp.de/value-sharing) with scope limited to a single CBOR data item, not across separate messages.

PoC

import cbor2
from io import BytesIO

# Message from trusted source containing a shareable value
msg1 = cbor2.dumps(cbor2.CBORTag(28, "secret"))

# Attacker-controlled message referencing index 0
msg2 = cbor2.dumps(cbor2.CBORTag(29, 0))

# Decoder reused across trust boundaries
decoder = cbor2.CBORDecoder(BytesIO(b''))
decoder.decode_from_bytes(msg1)
print(decoder.decode_from_bytes(msg2))  # prints "secret"

No special configuration required. Affects any application that reuses a CBORDecoder instance to decode messages from different sources.

Impact

Information disclosure. Applications that reuse a CBORDecoder across trust boundaries are vulnerable if the trusted messages use value sharing (tag 28) and an attacker can send messages containing shared references (tag 29). An attacker who can send a crafted CBOR message containing a sharedref tag can read values from previously decoded messages, potentially exposing sensitive data such as credentials, tokens, or private user data.

Related

A similar issue in the encoder could produce invalid CBOR with dangling shared references:

import cbor2
from io import BytesIO

# Create encoder with value sharing enabled
encoder = cbor2.CBOREncoder(BytesIO(), value_sharing=True)

# Persistent object that will be encoded multiple times
shared_obj = ['hello']

# First encode: array containing shared_obj twice
encoder.encode([shared_obj, shared_obj])
print(f'First encode: {encoder.fp.getvalue().hex()}')
# Output: d81c82d81c816568656c6c6fd81d01

# Second encode: just shared_obj
encoder.fp = BytesIO()
encoder.encode(shared_obj)
result = encoder.fp.getvalue()
print(f'Second encode: {result.hex()}')
# Output: d81d01  (just a shared reference to index 1!)

# Try to decode the second result as standalone CBOR
decoder = cbor2.CBORDecoder(BytesIO(result))
decoded = decoder.decode()
# FAILS: shared reference 1 not found

While primarily a correctness bug, it could cause denial of service if invalid CBOR is transmitted to downstream systems that fail to parse it, or cause silent data corruption if the dangling reference happens to resolve to an unrelated value.

It can also be considered a memory leak in both the decoder and encoder as references are held that will never be released as long as the decoder/encoder remains alive.

Suggested resolution

Add dedicated boolean flags to track when an encode/decode operation is in progress. Reset shared state only when the flag is False (top-level call). This ensures state is reset for standalone calls while preserving shared references for nested calls from hooks (which need access to the registry for cyclic structures).

Decoder (_decoding flag):

• decode(): set flag True, reset state, decode, set flag False
• decode_from_bytes(): reset state only when flag is False

Encoder (_encoding flag):

• encode(): set flag True, reset state, encode, set flag False
• encode_to_bytes(): reset state only when flag is False

Multiple vulnerabilities exist in cbor2 through version 5.7.0 in the decode_definite_long_string() function of the C extension decoder (source/decoder.c):

1. Integer Underflow Leading to Out-of-Bounds Read (CWE-191, CWE-125): An incorrect variable reference and missing state reset in the chunk processing loop causes buffer_length to not be reset to zero after UTF-8 character consumption. This results in subsequent chunk_length calculations producing negative values (e.g., chunk_length = 65536 - buffer_length), which are passed as signed integers to the read() method, potentially triggering unlimited read operations and resource exhaustion.
2. Memory Leak via Missing Reference Count Release (CWE-401): The main processing loop fails to release Python object references (Py_DECREF) for chunk objects allocated in each iteration. For CBOR strings longer than 65536 bytes, this causes cumulative memory leaks proportional to the payload size, enabling memory exhaustion attacks through repeated processing of large CBOR payloads.

Both vulnerabilities can be exploited remotely without authentication by sending specially-crafted CBOR data containing definite-length text strings with multi-byte UTF-8 characters positioned at 65536-byte chunk boundaries. Successful exploitation results in denial of service through process crashes (CBORDecodeEOF exceptions) or memory exhaustion. The vulnerabilities affect all applications using cbor2's C extension to process untrusted CBOR data, including web APIs, IoT data collectors, and message queue processors. Fixed in commit 851473490281f82d82560b2368284ef33cf6e8f9 pushed with released version 5.7.1.

Summary

• The cbor2 library is vulnerable to a Denial of Service (DoS) attack caused by uncontrolled recursion when decoding deeply nested CBOR structures.
• This vulnerability affects both the pure Python implementation and the C extension (_cbor2). The C extension correctly uses Python's C-API for recursion protection (Py_EnterRecursiveCall), but this mechanism is designed to prevent a stack overflow by raising a RecursionError. In some environments, this exception is not caught, thus causing the service process to terminate.
• While the library handles moderate nesting, it lacks a configurable, data-driven depth limit independent of Python's global recursion setting. An attacker can supply a crafted CBOR payload containing thousands of nested arrays (e.g., 0x81). When cbor2.loads() attempts to parse this, it hits the interpreter's recursion limit, causing the call to raise a RecursionError.
• By sending a stream of small (<100KB) malicious packets, an attacker can repeatedly crash worker processes faster than they can be restarted, resulting in a complete and sustained Denial of Service.

Details

• The vulnerability stems from the recursive design of the CBORDecoder class, specifically how it decodes nested container types like Arrays and Maps.
• Inside decode_array (and similarly decode_map), the decoder iterates through the number of elements specified in the CBOR header. For each element, it calls self.decode() again to parse the nested item. This recursive call lacks a depth-tracking mechanism.
• Vulnerable Code Locations:
  • cbor2/decoder.py (Pure Python implementation)
  • source/decoder.c (C extension implementation)
• Execution Flow:
  1. The cbor2.loads() function initializes a CBORDecoder and calls its decode() method.
  2. The decode() method reads the initial byte and dispatches control to a specific handler based on the major type. For an Array (Major Type 4), it calls decode_array.
  3. decode_array loops and calls self.decode() for each item, leading to deep recursion when parsing a payload like [...[...[1]...]...].

PoC

import cbor2

DEPTH = 1000

payload = b'\x81' * DEPTH + b'\x01'
print(f"[*] Payload size: {len(payload) / 1024:.2f} KB")
print("[*] Triggering decoder...")

try:
    cbor2.loads(payload)
    print("[+] Parsed successfully (Not Vulnerable)")
except RecursionError:
    print("\n[!] VULNERABLE: RecursionError triggered!")
except Exception as e:
    print(f"\n[-] Unexpected Error: {type(e).__name__}: {e}")

Impact

• Scope: This vulnerability affects any application using cbor2 to parse untrusted data. Common use cases include IoT data processing, WebAuthn (FIDO2) authentication flows, and inter-service communication over COSE (CBOR Object Signing and Encryption).
• Attack Vector: A remote, unauthenticated attacker can achieve a full Denial of Service with a highly efficient, low-bandwidth attack. A payload under 100KB is sufficient to reliably terminate a Python worker process.

Credit

This issue was discovered by Kevin Tu of TMIR at ByteDance. The patch was developed by @agronholm.

Summary

When a CBORDecoder instance is reused across multiple decode operations, values marked with the shareable tag (28) persist in memory and can be accessed by subsequent CBOR messages using the sharedref tag (29). This allows an attacker-controlled message to read data from previously decoded messages if the decoder is reused across trust boundaries.

Details

The issue is in the decoder's handling of the shareables list, which stores values tagged with CBOR tag 28 (shareable) for later reference by tag 29 (sharedref).

When decode_from_bytes() is called or when .fp is set to a new stream, the shareables list is not cleared. This allows references to persist across separate decode operations.

The issue exists in both the C extension and the pure Python decoder.

In the C extension (source/decoder.c), the _CBORDecoder_set_fp function (line ~202) updates the file pointer but does not reset the shareables state:

  static int
  _CBORDecoder_set_fp(CBORDecoderObject *self, PyObject *value, void *closure)
  {
      // ... validation ...
      tmp = self->read;
      self->read = read;
      Py_DECREF(tmp);
      return 0;
      // Missing: PyList_Clear(self->shareables) or equivalent
  }

In the pure Python decoder (cbor2/_decoder.py), the fp setter similarly fails to clear self._shareables.

Similarly, decode_from_bytes() in both implementations saves and restores the read pointer but does not clear the shareables list between decodes.

The shareable/sharedref tags are defined in the CBOR value sharing extension (http://cbor.schmorp.de/value-sharing) with scope limited to a single CBOR data item, not across separate messages.

PoC

import cbor2
from io import BytesIO

# Message from trusted source containing a shareable value
msg1 = cbor2.dumps(cbor2.CBORTag(28, "secret"))

# Attacker-controlled message referencing index 0
msg2 = cbor2.dumps(cbor2.CBORTag(29, 0))

# Decoder reused across trust boundaries
decoder = cbor2.CBORDecoder(BytesIO(b''))
decoder.decode_from_bytes(msg1)
print(decoder.decode_from_bytes(msg2))  # prints "secret"

No special configuration required. Affects any application that reuses a CBORDecoder instance to decode messages from different sources.

Impact

Information disclosure. Applications that reuse a CBORDecoder across trust boundaries are vulnerable if the trusted messages use value sharing (tag 28) and an attacker can send messages containing shared references (tag 29). An attacker who can send a crafted CBOR message containing a sharedref tag can read values from previously decoded messages, potentially exposing sensitive data such as credentials, tokens, or private user data.

Related

A similar issue in the encoder could produce invalid CBOR with dangling shared references:

import cbor2
from io import BytesIO

# Create encoder with value sharing enabled
encoder = cbor2.CBOREncoder(BytesIO(), value_sharing=True)

# Persistent object that will be encoded multiple times
shared_obj = ['hello']

# First encode: array containing shared_obj twice
encoder.encode([shared_obj, shared_obj])
print(f'First encode: {encoder.fp.getvalue().hex()}')
# Output: d81c82d81c816568656c6c6fd81d01

# Second encode: just shared_obj
encoder.fp = BytesIO()
encoder.encode(shared_obj)
result = encoder.fp.getvalue()
print(f'Second encode: {result.hex()}')
# Output: d81d01  (just a shared reference to index 1!)

# Try to decode the second result as standalone CBOR
decoder = cbor2.CBORDecoder(BytesIO(result))
decoded = decoder.decode()
# FAILS: shared reference 1 not found

While primarily a correctness bug, it could cause denial of service if invalid CBOR is transmitted to downstream systems that fail to parse it, or cause silent data corruption if the dangling reference happens to resolve to an unrelated value.

It can also be considered a memory leak in both the decoder and encoder as references are held that will never be released as long as the decoder/encoder remains alive.

Suggested resolution

Add dedicated boolean flags to track when an encode/decode operation is in progress. Reset shared state only when the flag is False (top-level call). This ensures state is reset for standalone calls while preserving shared references for nested calls from hooks (which need access to the registry for cyclic structures).

Decoder (_decoding flag):

• decode(): set flag True, reset state, decode, set flag False
• decode_from_bytes(): reset state only when flag is False

Encoder (_encoding flag):

• encode(): set flag True, reset state, encode, set flag False
• encode_to_bytes(): reset state only when flag is False

Multiple vulnerabilities exist in cbor2 through version 5.7.0 in the decode_definite_long_string() function of the C extension decoder (source/decoder.c):

1. Integer Underflow Leading to Out-of-Bounds Read (CWE-191, CWE-125): An incorrect variable reference and missing state reset in the chunk processing loop causes buffer_length to not be reset to zero after UTF-8 character consumption. This results in subsequent chunk_length calculations producing negative values (e.g., chunk_length = 65536 - buffer_length), which are passed as signed integers to the read() method, potentially triggering unlimited read operations and resource exhaustion.
2. Memory Leak via Missing Reference Count Release (CWE-401): The main processing loop fails to release Python object references (Py_DECREF) for chunk objects allocated in each iteration. For CBOR strings longer than 65536 bytes, this causes cumulative memory leaks proportional to the payload size, enabling memory exhaustion attacks through repeated processing of large CBOR payloads.

Both vulnerabilities can be exploited remotely without authentication by sending specially-crafted CBOR data containing definite-length text strings with multi-byte UTF-8 characters positioned at 65536-byte chunk boundaries. Successful exploitation results in denial of service through process crashes (CBORDecodeEOF exceptions) or memory exhaustion. The vulnerabilities affect all applications using cbor2's C extension to process untrusted CBOR data, including web APIs, IoT data collectors, and message queue processors. Fixed in commit 851473490281f82d82560b2368284ef33cf6e8f9 pushed with released version 5.7.1.

Summary

• The cbor2 library is vulnerable to a Denial of Service (DoS) attack caused by uncontrolled recursion when decoding deeply nested CBOR structures.
• This vulnerability affects both the pure Python implementation and the C extension (_cbor2). The C extension correctly uses Python's C-API for recursion protection (Py_EnterRecursiveCall), but this mechanism is designed to prevent a stack overflow by raising a RecursionError. In some environments, this exception is not caught, thus causing the service process to terminate.
• While the library handles moderate nesting, it lacks a configurable, data-driven depth limit independent of Python's global recursion setting. An attacker can supply a crafted CBOR payload containing thousands of nested arrays (e.g., 0x81). When cbor2.loads() attempts to parse this, it hits the interpreter's recursion limit, causing the call to raise a RecursionError.
• By sending a stream of small (<100KB) malicious packets, an attacker can repeatedly crash worker processes faster than they can be restarted, resulting in a complete and sustained Denial of Service.

Details

• The vulnerability stems from the recursive design of the CBORDecoder class, specifically how it decodes nested container types like Arrays and Maps.
• Inside decode_array (and similarly decode_map), the decoder iterates through the number of elements specified in the CBOR header. For each element, it calls self.decode() again to parse the nested item. This recursive call lacks a depth-tracking mechanism.
• Vulnerable Code Locations:
  • cbor2/decoder.py (Pure Python implementation)
  • source/decoder.c (C extension implementation)
• Execution Flow:
  1. The cbor2.loads() function initializes a CBORDecoder and calls its decode() method.
  2. The decode() method reads the initial byte and dispatches control to a specific handler based on the major type. For an Array (Major Type 4), it calls decode_array.
  3. decode_array loops and calls self.decode() for each item, leading to deep recursion when parsing a payload like [...[...[1]...]...].

PoC

import cbor2

DEPTH = 1000

payload = b'\x81' * DEPTH + b'\x01'
print(f"[*] Payload size: {len(payload) / 1024:.2f} KB")
print("[*] Triggering decoder...")

try:
    cbor2.loads(payload)
    print("[+] Parsed successfully (Not Vulnerable)")
except RecursionError:
    print("\n[!] VULNERABLE: RecursionError triggered!")
except Exception as e:
    print(f"\n[-] Unexpected Error: {type(e).__name__}: {e}")

Impact

• Scope: This vulnerability affects any application using cbor2 to parse untrusted data. Common use cases include IoT data processing, WebAuthn (FIDO2) authentication flows, and inter-service communication over COSE (CBOR Object Signing and Encryption).
• Attack Vector: A remote, unauthenticated attacker can achieve a full Denial of Service with a highly efficient, low-bandwidth attack. A payload under 100KB is sufficient to reliably terminate a Python worker process.

Credit

This issue was discovered by Kevin Tu of TMIR at ByteDance. The patch was developed by @agronholm.

Summary

When a CBORDecoder instance is reused across multiple decode operations, values marked with the shareable tag (28) persist in memory and can be accessed by subsequent CBOR messages using the sharedref tag (29). This allows an attacker-controlled message to read data from previously decoded messages if the decoder is reused across trust boundaries.

Details

The issue is in the decoder's handling of the shareables list, which stores values tagged with CBOR tag 28 (shareable) for later reference by tag 29 (sharedref).

When decode_from_bytes() is called or when .fp is set to a new stream, the shareables list is not cleared. This allows references to persist across separate decode operations.

The issue exists in both the C extension and the pure Python decoder.

In the C extension (source/decoder.c), the _CBORDecoder_set_fp function (line ~202) updates the file pointer but does not reset the shareables state:

  static int
  _CBORDecoder_set_fp(CBORDecoderObject *self, PyObject *value, void *closure)
  {
      // ... validation ...
      tmp = self->read;
      self->read = read;
      Py_DECREF(tmp);
      return 0;
      // Missing: PyList_Clear(self->shareables) or equivalent
  }

In the pure Python decoder (cbor2/_decoder.py), the fp setter similarly fails to clear self._shareables.

Similarly, decode_from_bytes() in both implementations saves and restores the read pointer but does not clear the shareables list between decodes.

The shareable/sharedref tags are defined in the CBOR value sharing extension (http://cbor.schmorp.de/value-sharing) with scope limited to a single CBOR data item, not across separate messages.

PoC

import cbor2
from io import BytesIO

# Message from trusted source containing a shareable value
msg1 = cbor2.dumps(cbor2.CBORTag(28, "secret"))

# Attacker-controlled message referencing index 0
msg2 = cbor2.dumps(cbor2.CBORTag(29, 0))

# Decoder reused across trust boundaries
decoder = cbor2.CBORDecoder(BytesIO(b''))
decoder.decode_from_bytes(msg1)
print(decoder.decode_from_bytes(msg2))  # prints "secret"

No special configuration required. Affects any application that reuses a CBORDecoder instance to decode messages from different sources.

Impact

Information disclosure. Applications that reuse a CBORDecoder across trust boundaries are vulnerable if the trusted messages use value sharing (tag 28) and an attacker can send messages containing shared references (tag 29). An attacker who can send a crafted CBOR message containing a sharedref tag can read values from previously decoded messages, potentially exposing sensitive data such as credentials, tokens, or private user data.

Related

A similar issue in the encoder could produce invalid CBOR with dangling shared references:

import cbor2
from io import BytesIO

# Create encoder with value sharing enabled
encoder = cbor2.CBOREncoder(BytesIO(), value_sharing=True)

# Persistent object that will be encoded multiple times
shared_obj = ['hello']

# First encode: array containing shared_obj twice
encoder.encode([shared_obj, shared_obj])
print(f'First encode: {encoder.fp.getvalue().hex()}')
# Output: d81c82d81c816568656c6c6fd81d01

# Second encode: just shared_obj
encoder.fp = BytesIO()
encoder.encode(shared_obj)
result = encoder.fp.getvalue()
print(f'Second encode: {result.hex()}')
# Output: d81d01  (just a shared reference to index 1!)

# Try to decode the second result as standalone CBOR
decoder = cbor2.CBORDecoder(BytesIO(result))
decoded = decoder.decode()
# FAILS: shared reference 1 not found

While primarily a correctness bug, it could cause denial of service if invalid CBOR is transmitted to downstream systems that fail to parse it, or cause silent data corruption if the dangling reference happens to resolve to an unrelated value.

It can also be considered a memory leak in both the decoder and encoder as references are held that will never be released as long as the decoder/encoder remains alive.

Suggested resolution

Add dedicated boolean flags to track when an encode/decode operation is in progress. Reset shared state only when the flag is False (top-level call). This ensures state is reset for standalone calls while preserving shared references for nested calls from hooks (which need access to the registry for cyclic structures).

Decoder (_decoding flag):

• decode(): set flag True, reset state, decode, set flag False
• decode_from_bytes(): reset state only when flag is False

Encoder (_encoding flag):

• encode(): set flag True, reset state, encode, set flag False
• encode_to_bytes(): reset state only when flag is False

Multiple vulnerabilities exist in cbor2 through version 5.7.0 in the decode_definite_long_string() function of the C extension decoder (source/decoder.c):

1. Integer Underflow Leading to Out-of-Bounds Read (CWE-191, CWE-125): An incorrect variable reference and missing state reset in the chunk processing loop causes buffer_length to not be reset to zero after UTF-8 character consumption. This results in subsequent chunk_length calculations producing negative values (e.g., chunk_length = 65536 - buffer_length), which are passed as signed integers to the read() method, potentially triggering unlimited read operations and resource exhaustion.
2. Memory Leak via Missing Reference Count Release (CWE-401): The main processing loop fails to release Python object references (Py_DECREF) for chunk objects allocated in each iteration. For CBOR strings longer than 65536 bytes, this causes cumulative memory leaks proportional to the payload size, enabling memory exhaustion attacks through repeated processing of large CBOR payloads.

Both vulnerabilities can be exploited remotely without authentication by sending specially-crafted CBOR data containing definite-length text strings with multi-byte UTF-8 characters positioned at 65536-byte chunk boundaries. Successful exploitation results in denial of service through process crashes (CBORDecodeEOF exceptions) or memory exhaustion. The vulnerabilities affect all applications using cbor2's C extension to process untrusted CBOR data, including web APIs, IoT data collectors, and message queue processors. Fixed in commit 851473490281f82d82560b2368284ef33cf6e8f9 pushed with released version 5.7.1.

Summary

• The cbor2 library is vulnerable to a Denial of Service (DoS) attack caused by uncontrolled recursion when decoding deeply nested CBOR structures.
• This vulnerability affects both the pure Python implementation and the C extension (_cbor2). The C extension correctly uses Python's C-API for recursion protection (Py_EnterRecursiveCall), but this mechanism is designed to prevent a stack overflow by raising a RecursionError. In some environments, this exception is not caught, thus causing the service process to terminate.
• While the library handles moderate nesting, it lacks a configurable, data-driven depth limit independent of Python's global recursion setting. An attacker can supply a crafted CBOR payload containing thousands of nested arrays (e.g., 0x81). When cbor2.loads() attempts to parse this, it hits the interpreter's recursion limit, causing the call to raise a RecursionError.
• By sending a stream of small (<100KB) malicious packets, an attacker can repeatedly crash worker processes faster than they can be restarted, resulting in a complete and sustained Denial of Service.

Details

• The vulnerability stems from the recursive design of the CBORDecoder class, specifically how it decodes nested container types like Arrays and Maps.
• Inside decode_array (and similarly decode_map), the decoder iterates through the number of elements specified in the CBOR header. For each element, it calls self.decode() again to parse the nested item. This recursive call lacks a depth-tracking mechanism.
• Vulnerable Code Locations:
  • cbor2/decoder.py (Pure Python implementation)
  • source/decoder.c (C extension implementation)
• Execution Flow:
  1. The cbor2.loads() function initializes a CBORDecoder and calls its decode() method.
  2. The decode() method reads the initial byte and dispatches control to a specific handler based on the major type. For an Array (Major Type 4), it calls decode_array.
  3. decode_array loops and calls self.decode() for each item, leading to deep recursion when parsing a payload like [...[...[1]...]...].

PoC

import cbor2

DEPTH = 1000

payload = b'\x81' * DEPTH + b'\x01'
print(f"[*] Payload size: {len(payload) / 1024:.2f} KB")
print("[*] Triggering decoder...")

try:
    cbor2.loads(payload)
    print("[+] Parsed successfully (Not Vulnerable)")
except RecursionError:
    print("\n[!] VULNERABLE: RecursionError triggered!")
except Exception as e:
    print(f"\n[-] Unexpected Error: {type(e).__name__}: {e}")

Impact

• Scope: This vulnerability affects any application using cbor2 to parse untrusted data. Common use cases include IoT data processing, WebAuthn (FIDO2) authentication flows, and inter-service communication over COSE (CBOR Object Signing and Encryption).
• Attack Vector: A remote, unauthenticated attacker can achieve a full Denial of Service with a highly efficient, low-bandwidth attack. A payload under 100KB is sufficient to reliably terminate a Python worker process.

Credit

This issue was discovered by Kevin Tu of TMIR at ByteDance. The patch was developed by @agronholm.

Summary

When a CBORDecoder instance is reused across multiple decode operations, values marked with the shareable tag (28) persist in memory and can be accessed by subsequent CBOR messages using the sharedref tag (29). This allows an attacker-controlled message to read data from previously decoded messages if the decoder is reused across trust boundaries.

Details

The issue is in the decoder's handling of the shareables list, which stores values tagged with CBOR tag 28 (shareable) for later reference by tag 29 (sharedref).

When decode_from_bytes() is called or when .fp is set to a new stream, the shareables list is not cleared. This allows references to persist across separate decode operations.

The issue exists in both the C extension and the pure Python decoder.

In the C extension (source/decoder.c), the _CBORDecoder_set_fp function (line ~202) updates the file pointer but does not reset the shareables state:

  static int
  _CBORDecoder_set_fp(CBORDecoderObject *self, PyObject *value, void *closure)
  {
      // ... validation ...
      tmp = self->read;
      self->read = read;
      Py_DECREF(tmp);
      return 0;
      // Missing: PyList_Clear(self->shareables) or equivalent
  }

In the pure Python decoder (cbor2/_decoder.py), the fp setter similarly fails to clear self._shareables.

Similarly, decode_from_bytes() in both implementations saves and restores the read pointer but does not clear the shareables list between decodes.

The shareable/sharedref tags are defined in the CBOR value sharing extension (http://cbor.schmorp.de/value-sharing) with scope limited to a single CBOR data item, not across separate messages.

PoC

import cbor2
from io import BytesIO

# Message from trusted source containing a shareable value
msg1 = cbor2.dumps(cbor2.CBORTag(28, "secret"))

# Attacker-controlled message referencing index 0
msg2 = cbor2.dumps(cbor2.CBORTag(29, 0))

# Decoder reused across trust boundaries
decoder = cbor2.CBORDecoder(BytesIO(b''))
decoder.decode_from_bytes(msg1)
print(decoder.decode_from_bytes(msg2))  # prints "secret"

No special configuration required. Affects any application that reuses a CBORDecoder instance to decode messages from different sources.

Impact

Information disclosure. Applications that reuse a CBORDecoder across trust boundaries are vulnerable if the trusted messages use value sharing (tag 28) and an attacker can send messages containing shared references (tag 29). An attacker who can send a crafted CBOR message containing a sharedref tag can read values from previously decoded messages, potentially exposing sensitive data such as credentials, tokens, or private user data.

Related

A similar issue in the encoder could produce invalid CBOR with dangling shared references:

import cbor2
from io import BytesIO

# Create encoder with value sharing enabled
encoder = cbor2.CBOREncoder(BytesIO(), value_sharing=True)

# Persistent object that will be encoded multiple times
shared_obj = ['hello']

# First encode: array containing shared_obj twice
encoder.encode([shared_obj, shared_obj])
print(f'First encode: {encoder.fp.getvalue().hex()}')
# Output: d81c82d81c816568656c6c6fd81d01

# Second encode: just shared_obj
encoder.fp = BytesIO()
encoder.encode(shared_obj)
result = encoder.fp.getvalue()
print(f'Second encode: {result.hex()}')
# Output: d81d01  (just a shared reference to index 1!)

# Try to decode the second result as standalone CBOR
decoder = cbor2.CBORDecoder(BytesIO(result))
decoded = decoder.decode()
# FAILS: shared reference 1 not found

While primarily a correctness bug, it could cause denial of service if invalid CBOR is transmitted to downstream systems that fail to parse it, or cause silent data corruption if the dangling reference happens to resolve to an unrelated value.

It can also be considered a memory leak in both the decoder and encoder as references are held that will never be released as long as the decoder/encoder remains alive.

Suggested resolution

Add dedicated boolean flags to track when an encode/decode operation is in progress. Reset shared state only when the flag is False (top-level call). This ensures state is reset for standalone calls while preserving shared references for nested calls from hooks (which need access to the registry for cyclic structures).

Decoder (_decoding flag):

• decode(): set flag True, reset state, decode, set flag False
• decode_from_bytes(): reset state only when flag is False

Encoder (_encoding flag):

• encode(): set flag True, reset state, encode, set flag False
• encode_to_bytes(): reset state only when flag is False

Multiple vulnerabilities exist in cbor2 through version 5.7.0 in the decode_definite_long_string() function of the C extension decoder (source/decoder.c):

1. Integer Underflow Leading to Out-of-Bounds Read (CWE-191, CWE-125): An incorrect variable reference and missing state reset in the chunk processing loop causes buffer_length to not be reset to zero after UTF-8 character consumption. This results in subsequent chunk_length calculations producing negative values (e.g., chunk_length = 65536 - buffer_length), which are passed as signed integers to the read() method, potentially triggering unlimited read operations and resource exhaustion.
2. Memory Leak via Missing Reference Count Release (CWE-401): The main processing loop fails to release Python object references (Py_DECREF) for chunk objects allocated in each iteration. For CBOR strings longer than 65536 bytes, this causes cumulative memory leaks proportional to the payload size, enabling memory exhaustion attacks through repeated processing of large CBOR payloads.

Both vulnerabilities can be exploited remotely without authentication by sending specially-crafted CBOR data containing definite-length text strings with multi-byte UTF-8 characters positioned at 65536-byte chunk boundaries. Successful exploitation results in denial of service through process crashes (CBORDecodeEOF exceptions) or memory exhaustion. The vulnerabilities affect all applications using cbor2's C extension to process untrusted CBOR data, including web APIs, IoT data collectors, and message queue processors. Fixed in commit 851473490281f82d82560b2368284ef33cf6e8f9 pushed with released version 5.7.1.

Summary

Ever since https://github.com/agronholm/cbor2/pull/204 (or specifically https://github.com/agronholm/cbor2/commit/387755eacf0be35591a478d3c67fe10618a6d542) was merged, I can create a reproducible crash when running the snippet under PoC on a current Debian bullseye aarm64 on a Raspberry Pi 3 (I was not able to reproduce this on my x86_64 Laptop with Python 3.11; I suspect because there is enough memory to allocate still)

Details

PoC

import json
import concurrent.futures
import cbor2

def test():
    obj = "x" * 131128
    cbor_enc = cbor2.dumps(obj)
    return cbor2.loads(cbor_enc)

with concurrent.futures.ProcessPoolExecutor() as executor:
    future = executor.submit(test)
    print(future.result())

malloc(): unsorted double linked list corrupted
Traceback (most recent call last):
  File "test.py", line 14, in <module>
    print(future.result())
  File "/usr/lib/python3.9/concurrent/futures/_base.py", line 440, in result
    return self.__get_result()
  File "/usr/lib/python3.9/concurrent/futures/_base.py", line 389, in __get_result
    raise self._exception
concurrent.futures.process.BrokenProcessPool: A process in the process pool was terminated abruptly while the future was running or pending.

If one calls it without the indirection via the pool executor, a SystemError is shown that hides the buffer overflow.

import json
import cbor2

def test():
    obj = "x" * 131128
    cbor_enc = cbor2.dumps(obj)
    return cbor2.loads(cbor_enc)

print(test())

Traceback (most recent call last):
  File "test.py", line 12, in <module>
    print(test())
  File "test.py", line 9, in test
    return cbor2.loads(cbor_enc)
SystemError: <built-in function loads> returned NULL without setting an error

Impact

An attacker can crash a service using cbor2 to parse a CBOR binary by sending a long enough object.

Summary

• The cbor2 library is vulnerable to a Denial of Service (DoS) attack caused by uncontrolled recursion when decoding deeply nested CBOR structures.
• This vulnerability affects both the pure Python implementation and the C extension (_cbor2). The C extension correctly uses Python's C-API for recursion protection (Py_EnterRecursiveCall), but this mechanism is designed to prevent a stack overflow by raising a RecursionError. In some environments, this exception is not caught, thus causing the service process to terminate.
• While the library handles moderate nesting, it lacks a configurable, data-driven depth limit independent of Python's global recursion setting. An attacker can supply a crafted CBOR payload containing thousands of nested arrays (e.g., 0x81). When cbor2.loads() attempts to parse this, it hits the interpreter's recursion limit, causing the call to raise a RecursionError.
• By sending a stream of small (<100KB) malicious packets, an attacker can repeatedly crash worker processes faster than they can be restarted, resulting in a complete and sustained Denial of Service.

Details

• The vulnerability stems from the recursive design of the CBORDecoder class, specifically how it decodes nested container types like Arrays and Maps.
• Inside decode_array (and similarly decode_map), the decoder iterates through the number of elements specified in the CBOR header. For each element, it calls self.decode() again to parse the nested item. This recursive call lacks a depth-tracking mechanism.
• Vulnerable Code Locations:
  • cbor2/decoder.py (Pure Python implementation)
  • source/decoder.c (C extension implementation)
• Execution Flow:
  1. The cbor2.loads() function initializes a CBORDecoder and calls its decode() method.
  2. The decode() method reads the initial byte and dispatches control to a specific handler based on the major type. For an Array (Major Type 4), it calls decode_array.
  3. decode_array loops and calls self.decode() for each item, leading to deep recursion when parsing a payload like [...[...[1]...]...].

PoC

import cbor2

DEPTH = 1000

payload = b'\x81' * DEPTH + b'\x01'
print(f"[*] Payload size: {len(payload) / 1024:.2f} KB")
print("[*] Triggering decoder...")

try:
    cbor2.loads(payload)
    print("[+] Parsed successfully (Not Vulnerable)")
except RecursionError:
    print("\n[!] VULNERABLE: RecursionError triggered!")
except Exception as e:
    print(f"\n[-] Unexpected Error: {type(e).__name__}: {e}")

Impact

• Scope: This vulnerability affects any application using cbor2 to parse untrusted data. Common use cases include IoT data processing, WebAuthn (FIDO2) authentication flows, and inter-service communication over COSE (CBOR Object Signing and Encryption).
• Attack Vector: A remote, unauthenticated attacker can achieve a full Denial of Service with a highly efficient, low-bandwidth attack. A payload under 100KB is sufficient to reliably terminate a Python worker process.

Credit

This issue was discovered by Kevin Tu of TMIR at ByteDance. The patch was developed by @agronholm.

Summary

When a CBORDecoder instance is reused across multiple decode operations, values marked with the shareable tag (28) persist in memory and can be accessed by subsequent CBOR messages using the sharedref tag (29). This allows an attacker-controlled message to read data from previously decoded messages if the decoder is reused across trust boundaries.

Details

The issue is in the decoder's handling of the shareables list, which stores values tagged with CBOR tag 28 (shareable) for later reference by tag 29 (sharedref).

When decode_from_bytes() is called or when .fp is set to a new stream, the shareables list is not cleared. This allows references to persist across separate decode operations.

The issue exists in both the C extension and the pure Python decoder.

In the C extension (source/decoder.c), the _CBORDecoder_set_fp function (line ~202) updates the file pointer but does not reset the shareables state:

  static int
  _CBORDecoder_set_fp(CBORDecoderObject *self, PyObject *value, void *closure)
  {
      // ... validation ...
      tmp = self->read;
      self->read = read;
      Py_DECREF(tmp);
      return 0;
      // Missing: PyList_Clear(self->shareables) or equivalent
  }

In the pure Python decoder (cbor2/_decoder.py), the fp setter similarly fails to clear self._shareables.

Similarly, decode_from_bytes() in both implementations saves and restores the read pointer but does not clear the shareables list between decodes.

The shareable/sharedref tags are defined in the CBOR value sharing extension (http://cbor.schmorp.de/value-sharing) with scope limited to a single CBOR data item, not across separate messages.

PoC

import cbor2
from io import BytesIO

# Message from trusted source containing a shareable value
msg1 = cbor2.dumps(cbor2.CBORTag(28, "secret"))

# Attacker-controlled message referencing index 0
msg2 = cbor2.dumps(cbor2.CBORTag(29, 0))

# Decoder reused across trust boundaries
decoder = cbor2.CBORDecoder(BytesIO(b''))
decoder.decode_from_bytes(msg1)
print(decoder.decode_from_bytes(msg2))  # prints "secret"

No special configuration required. Affects any application that reuses a CBORDecoder instance to decode messages from different sources.

Impact

Information disclosure. Applications that reuse a CBORDecoder across trust boundaries are vulnerable if the trusted messages use value sharing (tag 28) and an attacker can send messages containing shared references (tag 29). An attacker who can send a crafted CBOR message containing a sharedref tag can read values from previously decoded messages, potentially exposing sensitive data such as credentials, tokens, or private user data.

Related

A similar issue in the encoder could produce invalid CBOR with dangling shared references:

import cbor2
from io import BytesIO

# Create encoder with value sharing enabled
encoder = cbor2.CBOREncoder(BytesIO(), value_sharing=True)

# Persistent object that will be encoded multiple times
shared_obj = ['hello']

# First encode: array containing shared_obj twice
encoder.encode([shared_obj, shared_obj])
print(f'First encode: {encoder.fp.getvalue().hex()}')
# Output: d81c82d81c816568656c6c6fd81d01

# Second encode: just shared_obj
encoder.fp = BytesIO()
encoder.encode(shared_obj)
result = encoder.fp.getvalue()
print(f'Second encode: {result.hex()}')
# Output: d81d01  (just a shared reference to index 1!)

# Try to decode the second result as standalone CBOR
decoder = cbor2.CBORDecoder(BytesIO(result))
decoded = decoder.decode()
# FAILS: shared reference 1 not found

While primarily a correctness bug, it could cause denial of service if invalid CBOR is transmitted to downstream systems that fail to parse it, or cause silent data corruption if the dangling reference happens to resolve to an unrelated value.

It can also be considered a memory leak in both the decoder and encoder as references are held that will never be released as long as the decoder/encoder remains alive.

Suggested resolution

Add dedicated boolean flags to track when an encode/decode operation is in progress. Reset shared state only when the flag is False (top-level call). This ensures state is reset for standalone calls while preserving shared references for nested calls from hooks (which need access to the registry for cyclic structures).

Decoder (_decoding flag):

• decode(): set flag True, reset state, decode, set flag False
• decode_from_bytes(): reset state only when flag is False

Encoder (_encoding flag):

• encode(): set flag True, reset state, encode, set flag False
• encode_to_bytes(): reset state only when flag is False

Multiple vulnerabilities exist in cbor2 through version 5.7.0 in the decode_definite_long_string() function of the C extension decoder (source/decoder.c):

1. Integer Underflow Leading to Out-of-Bounds Read (CWE-191, CWE-125): An incorrect variable reference and missing state reset in the chunk processing loop causes buffer_length to not be reset to zero after UTF-8 character consumption. This results in subsequent chunk_length calculations producing negative values (e.g., chunk_length = 65536 - buffer_length), which are passed as signed integers to the read() method, potentially triggering unlimited read operations and resource exhaustion.
2. Memory Leak via Missing Reference Count Release (CWE-401): The main processing loop fails to release Python object references (Py_DECREF) for chunk objects allocated in each iteration. For CBOR strings longer than 65536 bytes, this causes cumulative memory leaks proportional to the payload size, enabling memory exhaustion attacks through repeated processing of large CBOR payloads.

Both vulnerabilities can be exploited remotely without authentication by sending specially-crafted CBOR data containing definite-length text strings with multi-byte UTF-8 characters positioned at 65536-byte chunk boundaries. Successful exploitation results in denial of service through process crashes (CBORDecodeEOF exceptions) or memory exhaustion. The vulnerabilities affect all applications using cbor2's C extension to process untrusted CBOR data, including web APIs, IoT data collectors, and message queue processors. Fixed in commit 851473490281f82d82560b2368284ef33cf6e8f9 pushed with released version 5.7.1.

Summary

Ever since https://github.com/agronholm/cbor2/pull/204 (or specifically https://github.com/agronholm/cbor2/commit/387755eacf0be35591a478d3c67fe10618a6d542) was merged, I can create a reproducible crash when running the snippet under PoC on a current Debian bullseye aarm64 on a Raspberry Pi 3 (I was not able to reproduce this on my x86_64 Laptop with Python 3.11; I suspect because there is enough memory to allocate still)

Details

PoC

import json
import concurrent.futures
import cbor2

def test():
    obj = "x" * 131128
    cbor_enc = cbor2.dumps(obj)
    return cbor2.loads(cbor_enc)

with concurrent.futures.ProcessPoolExecutor() as executor:
    future = executor.submit(test)
    print(future.result())

malloc(): unsorted double linked list corrupted
Traceback (most recent call last):
  File "test.py", line 14, in <module>
    print(future.result())
  File "/usr/lib/python3.9/concurrent/futures/_base.py", line 440, in result
    return self.__get_result()
  File "/usr/lib/python3.9/concurrent/futures/_base.py", line 389, in __get_result
    raise self._exception
concurrent.futures.process.BrokenProcessPool: A process in the process pool was terminated abruptly while the future was running or pending.

If one calls it without the indirection via the pool executor, a SystemError is shown that hides the buffer overflow.

import json
import cbor2

def test():
    obj = "x" * 131128
    cbor_enc = cbor2.dumps(obj)
    return cbor2.loads(cbor_enc)

print(test())

Traceback (most recent call last):
  File "test.py", line 12, in <module>
    print(test())
  File "test.py", line 9, in test
    return cbor2.loads(cbor_enc)
SystemError: <built-in function loads> returned NULL without setting an error

Impact

An attacker can crash a service using cbor2 to parse a CBOR binary by sending a long enough object.

Summary

• The cbor2 library is vulnerable to a Denial of Service (DoS) attack caused by uncontrolled recursion when decoding deeply nested CBOR structures.
• This vulnerability affects both the pure Python implementation and the C extension (_cbor2). The C extension correctly uses Python's C-API for recursion protection (Py_EnterRecursiveCall), but this mechanism is designed to prevent a stack overflow by raising a RecursionError. In some environments, this exception is not caught, thus causing the service process to terminate.
• While the library handles moderate nesting, it lacks a configurable, data-driven depth limit independent of Python's global recursion setting. An attacker can supply a crafted CBOR payload containing thousands of nested arrays (e.g., 0x81). When cbor2.loads() attempts to parse this, it hits the interpreter's recursion limit, causing the call to raise a RecursionError.
• By sending a stream of small (<100KB) malicious packets, an attacker can repeatedly crash worker processes faster than they can be restarted, resulting in a complete and sustained Denial of Service.

Details

• The vulnerability stems from the recursive design of the CBORDecoder class, specifically how it decodes nested container types like Arrays and Maps.
• Inside decode_array (and similarly decode_map), the decoder iterates through the number of elements specified in the CBOR header. For each element, it calls self.decode() again to parse the nested item. This recursive call lacks a depth-tracking mechanism.
• Vulnerable Code Locations:
  • cbor2/decoder.py (Pure Python implementation)
  • source/decoder.c (C extension implementation)
• Execution Flow:
  1. The cbor2.loads() function initializes a CBORDecoder and calls its decode() method.
  2. The decode() method reads the initial byte and dispatches control to a specific handler based on the major type. For an Array (Major Type 4), it calls decode_array.
  3. decode_array loops and calls self.decode() for each item, leading to deep recursion when parsing a payload like [...[...[1]...]...].

PoC

import cbor2

DEPTH = 1000

payload = b'\x81' * DEPTH + b'\x01'
print(f"[*] Payload size: {len(payload) / 1024:.2f} KB")
print("[*] Triggering decoder...")

try:
    cbor2.loads(payload)
    print("[+] Parsed successfully (Not Vulnerable)")
except RecursionError:
    print("\n[!] VULNERABLE: RecursionError triggered!")
except Exception as e:
    print(f"\n[-] Unexpected Error: {type(e).__name__}: {e}")

Impact

• Scope: This vulnerability affects any application using cbor2 to parse untrusted data. Common use cases include IoT data processing, WebAuthn (FIDO2) authentication flows, and inter-service communication over COSE (CBOR Object Signing and Encryption).
• Attack Vector: A remote, unauthenticated attacker can achieve a full Denial of Service with a highly efficient, low-bandwidth attack. A payload under 100KB is sufficient to reliably terminate a Python worker process.

Credit

This issue was discovered by Kevin Tu of TMIR at ByteDance. The patch was developed by @agronholm.

Summary

When a CBORDecoder instance is reused across multiple decode operations, values marked with the shareable tag (28) persist in memory and can be accessed by subsequent CBOR messages using the sharedref tag (29). This allows an attacker-controlled message to read data from previously decoded messages if the decoder is reused across trust boundaries.

Details

The issue is in the decoder's handling of the shareables list, which stores values tagged with CBOR tag 28 (shareable) for later reference by tag 29 (sharedref).

When decode_from_bytes() is called or when .fp is set to a new stream, the shareables list is not cleared. This allows references to persist across separate decode operations.

The issue exists in both the C extension and the pure Python decoder.

In the C extension (source/decoder.c), the _CBORDecoder_set_fp function (line ~202) updates the file pointer but does not reset the shareables state:

  static int
  _CBORDecoder_set_fp(CBORDecoderObject *self, PyObject *value, void *closure)
  {
      // ... validation ...
      tmp = self->read;
      self->read = read;
      Py_DECREF(tmp);
      return 0;
      // Missing: PyList_Clear(self->shareables) or equivalent
  }

In the pure Python decoder (cbor2/_decoder.py), the fp setter similarly fails to clear self._shareables.

Similarly, decode_from_bytes() in both implementations saves and restores the read pointer but does not clear the shareables list between decodes.

The shareable/sharedref tags are defined in the CBOR value sharing extension (http://cbor.schmorp.de/value-sharing) with scope limited to a single CBOR data item, not across separate messages.

PoC

import cbor2
from io import BytesIO

# Message from trusted source containing a shareable value
msg1 = cbor2.dumps(cbor2.CBORTag(28, "secret"))

# Attacker-controlled message referencing index 0
msg2 = cbor2.dumps(cbor2.CBORTag(29, 0))

# Decoder reused across trust boundaries
decoder = cbor2.CBORDecoder(BytesIO(b''))
decoder.decode_from_bytes(msg1)
print(decoder.decode_from_bytes(msg2))  # prints "secret"

No special configuration required. Affects any application that reuses a CBORDecoder instance to decode messages from different sources.

Impact

Information disclosure. Applications that reuse a CBORDecoder across trust boundaries are vulnerable if the trusted messages use value sharing (tag 28) and an attacker can send messages containing shared references (tag 29). An attacker who can send a crafted CBOR message containing a sharedref tag can read values from previously decoded messages, potentially exposing sensitive data such as credentials, tokens, or private user data.

Related

A similar issue in the encoder could produce invalid CBOR with dangling shared references:

import cbor2
from io import BytesIO

# Create encoder with value sharing enabled
encoder = cbor2.CBOREncoder(BytesIO(), value_sharing=True)

# Persistent object that will be encoded multiple times
shared_obj = ['hello']

# First encode: array containing shared_obj twice
encoder.encode([shared_obj, shared_obj])
print(f'First encode: {encoder.fp.getvalue().hex()}')
# Output: d81c82d81c816568656c6c6fd81d01

# Second encode: just shared_obj
encoder.fp = BytesIO()
encoder.encode(shared_obj)
result = encoder.fp.getvalue()
print(f'Second encode: {result.hex()}')
# Output: d81d01  (just a shared reference to index 1!)

# Try to decode the second result as standalone CBOR
decoder = cbor2.CBORDecoder(BytesIO(result))
decoded = decoder.decode()
# FAILS: shared reference 1 not found

While primarily a correctness bug, it could cause denial of service if invalid CBOR is transmitted to downstream systems that fail to parse it, or cause silent data corruption if the dangling reference happens to resolve to an unrelated value.

It can also be considered a memory leak in both the decoder and encoder as references are held that will never be released as long as the decoder/encoder remains alive.

Suggested resolution

Add dedicated boolean flags to track when an encode/decode operation is in progress. Reset shared state only when the flag is False (top-level call). This ensures state is reset for standalone calls while preserving shared references for nested calls from hooks (which need access to the registry for cyclic structures).

Decoder (_decoding flag):

• decode(): set flag True, reset state, decode, set flag False
• decode_from_bytes(): reset state only when flag is False

Encoder (_encoding flag):

• encode(): set flag True, reset state, encode, set flag False
• encode_to_bytes(): reset state only when flag is False

Multiple vulnerabilities exist in cbor2 through version 5.7.0 in the decode_definite_long_string() function of the C extension decoder (source/decoder.c):

1. Integer Underflow Leading to Out-of-Bounds Read (CWE-191, CWE-125): An incorrect variable reference and missing state reset in the chunk processing loop causes buffer_length to not be reset to zero after UTF-8 character consumption. This results in subsequent chunk_length calculations producing negative values (e.g., chunk_length = 65536 - buffer_length), which are passed as signed integers to the read() method, potentially triggering unlimited read operations and resource exhaustion.
2. Memory Leak via Missing Reference Count Release (CWE-401): The main processing loop fails to release Python object references (Py_DECREF) for chunk objects allocated in each iteration. For CBOR strings longer than 65536 bytes, this causes cumulative memory leaks proportional to the payload size, enabling memory exhaustion attacks through repeated processing of large CBOR payloads.

Both vulnerabilities can be exploited remotely without authentication by sending specially-crafted CBOR data containing definite-length text strings with multi-byte UTF-8 characters positioned at 65536-byte chunk boundaries. Successful exploitation results in denial of service through process crashes (CBORDecodeEOF exceptions) or memory exhaustion. The vulnerabilities affect all applications using cbor2's C extension to process untrusted CBOR data, including web APIs, IoT data collectors, and message queue processors. Fixed in commit 851473490281f82d82560b2368284ef33cf6e8f9 pushed with released version 5.7.1.

Summary

Ever since https://github.com/agronholm/cbor2/pull/204 (or specifically https://github.com/agronholm/cbor2/commit/387755eacf0be35591a478d3c67fe10618a6d542) was merged, I can create a reproducible crash when running the snippet under PoC on a current Debian bullseye aarm64 on a Raspberry Pi 3 (I was not able to reproduce this on my x86_64 Laptop with Python 3.11; I suspect because there is enough memory to allocate still)

Details

PoC

import json
import concurrent.futures
import cbor2

def test():
    obj = "x" * 131128
    cbor_enc = cbor2.dumps(obj)
    return cbor2.loads(cbor_enc)

with concurrent.futures.ProcessPoolExecutor() as executor:
    future = executor.submit(test)
    print(future.result())

malloc(): unsorted double linked list corrupted
Traceback (most recent call last):
  File "test.py", line 14, in <module>
    print(future.result())
  File "/usr/lib/python3.9/concurrent/futures/_base.py", line 440, in result
    return self.__get_result()
  File "/usr/lib/python3.9/concurrent/futures/_base.py", line 389, in __get_result
    raise self._exception
concurrent.futures.process.BrokenProcessPool: A process in the process pool was terminated abruptly while the future was running or pending.

If one calls it without the indirection via the pool executor, a SystemError is shown that hides the buffer overflow.

import json
import cbor2

def test():
    obj = "x" * 131128
    cbor_enc = cbor2.dumps(obj)
    return cbor2.loads(cbor_enc)

print(test())

Traceback (most recent call last):
  File "test.py", line 12, in <module>
    print(test())
  File "test.py", line 9, in test
    return cbor2.loads(cbor_enc)
SystemError: <built-in function loads> returned NULL without setting an error

Impact

An attacker can crash a service using cbor2 to parse a CBOR binary by sending a long enough object.

Summary

• The cbor2 library is vulnerable to a Denial of Service (DoS) attack caused by uncontrolled recursion when decoding deeply nested CBOR structures.
• This vulnerability affects both the pure Python implementation and the C extension (_cbor2). The C extension correctly uses Python's C-API for recursion protection (Py_EnterRecursiveCall), but this mechanism is designed to prevent a stack overflow by raising a RecursionError. In some environments, this exception is not caught, thus causing the service process to terminate.
• While the library handles moderate nesting, it lacks a configurable, data-driven depth limit independent of Python's global recursion setting. An attacker can supply a crafted CBOR payload containing thousands of nested arrays (e.g., 0x81). When cbor2.loads() attempts to parse this, it hits the interpreter's recursion limit, causing the call to raise a RecursionError.
• By sending a stream of small (<100KB) malicious packets, an attacker can repeatedly crash worker processes faster than they can be restarted, resulting in a complete and sustained Denial of Service.

Details

• The vulnerability stems from the recursive design of the CBORDecoder class, specifically how it decodes nested container types like Arrays and Maps.
• Inside decode_array (and similarly decode_map), the decoder iterates through the number of elements specified in the CBOR header. For each element, it calls self.decode() again to parse the nested item. This recursive call lacks a depth-tracking mechanism.
• Vulnerable Code Locations:
  • cbor2/decoder.py (Pure Python implementation)
  • source/decoder.c (C extension implementation)
• Execution Flow:
  1. The cbor2.loads() function initializes a CBORDecoder and calls its decode() method.
  2. The decode() method reads the initial byte and dispatches control to a specific handler based on the major type. For an Array (Major Type 4), it calls decode_array.
  3. decode_array loops and calls self.decode() for each item, leading to deep recursion when parsing a payload like [...[...[1]...]...].

PoC

import cbor2

DEPTH = 1000

payload = b'\x81' * DEPTH + b'\x01'
print(f"[*] Payload size: {len(payload) / 1024:.2f} KB")
print("[*] Triggering decoder...")

try:
    cbor2.loads(payload)
    print("[+] Parsed successfully (Not Vulnerable)")
except RecursionError:
    print("\n[!] VULNERABLE: RecursionError triggered!")
except Exception as e:
    print(f"\n[-] Unexpected Error: {type(e).__name__}: {e}")

Impact

• Scope: This vulnerability affects any application using cbor2 to parse untrusted data. Common use cases include IoT data processing, WebAuthn (FIDO2) authentication flows, and inter-service communication over COSE (CBOR Object Signing and Encryption).
• Attack Vector: A remote, unauthenticated attacker can achieve a full Denial of Service with a highly efficient, low-bandwidth attack. A payload under 100KB is sufficient to reliably terminate a Python worker process.

Credit

This issue was discovered by Kevin Tu of TMIR at ByteDance. The patch was developed by @agronholm.

Summary

When a CBORDecoder instance is reused across multiple decode operations, values marked with the shareable tag (28) persist in memory and can be accessed by subsequent CBOR messages using the sharedref tag (29). This allows an attacker-controlled message to read data from previously decoded messages if the decoder is reused across trust boundaries.

Details

The issue is in the decoder's handling of the shareables list, which stores values tagged with CBOR tag 28 (shareable) for later reference by tag 29 (sharedref).

When decode_from_bytes() is called or when .fp is set to a new stream, the shareables list is not cleared. This allows references to persist across separate decode operations.

The issue exists in both the C extension and the pure Python decoder.

In the C extension (source/decoder.c), the _CBORDecoder_set_fp function (line ~202) updates the file pointer but does not reset the shareables state:

  static int
  _CBORDecoder_set_fp(CBORDecoderObject *self, PyObject *value, void *closure)
  {
      // ... validation ...
      tmp = self->read;
      self->read = read;
      Py_DECREF(tmp);
      return 0;
      // Missing: PyList_Clear(self->shareables) or equivalent
  }

In the pure Python decoder (cbor2/_decoder.py), the fp setter similarly fails to clear self._shareables.

Similarly, decode_from_bytes() in both implementations saves and restores the read pointer but does not clear the shareables list between decodes.

The shareable/sharedref tags are defined in the CBOR value sharing extension (http://cbor.schmorp.de/value-sharing) with scope limited to a single CBOR data item, not across separate messages.

PoC

import cbor2
from io import BytesIO

# Message from trusted source containing a shareable value
msg1 = cbor2.dumps(cbor2.CBORTag(28, "secret"))

# Attacker-controlled message referencing index 0
msg2 = cbor2.dumps(cbor2.CBORTag(29, 0))

# Decoder reused across trust boundaries
decoder = cbor2.CBORDecoder(BytesIO(b''))
decoder.decode_from_bytes(msg1)
print(decoder.decode_from_bytes(msg2))  # prints "secret"

No special configuration required. Affects any application that reuses a CBORDecoder instance to decode messages from different sources.

Impact

Information disclosure. Applications that reuse a CBORDecoder across trust boundaries are vulnerable if the trusted messages use value sharing (tag 28) and an attacker can send messages containing shared references (tag 29). An attacker who can send a crafted CBOR message containing a sharedref tag can read values from previously decoded messages, potentially exposing sensitive data such as credentials, tokens, or private user data.

Related

A similar issue in the encoder could produce invalid CBOR with dangling shared references:

import cbor2
from io import BytesIO

# Create encoder with value sharing enabled
encoder = cbor2.CBOREncoder(BytesIO(), value_sharing=True)

# Persistent object that will be encoded multiple times
shared_obj = ['hello']

# First encode: array containing shared_obj twice
encoder.encode([shared_obj, shared_obj])
print(f'First encode: {encoder.fp.getvalue().hex()}')
# Output: d81c82d81c816568656c6c6fd81d01

# Second encode: just shared_obj
encoder.fp = BytesIO()
encoder.encode(shared_obj)
result = encoder.fp.getvalue()
print(f'Second encode: {result.hex()}')
# Output: d81d01  (just a shared reference to index 1!)

# Try to decode the second result as standalone CBOR
decoder = cbor2.CBORDecoder(BytesIO(result))
decoded = decoder.decode()
# FAILS: shared reference 1 not found

While primarily a correctness bug, it could cause denial of service if invalid CBOR is transmitted to downstream systems that fail to parse it, or cause silent data corruption if the dangling reference happens to resolve to an unrelated value.

It can also be considered a memory leak in both the decoder and encoder as references are held that will never be released as long as the decoder/encoder remains alive.

Suggested resolution

Add dedicated boolean flags to track when an encode/decode operation is in progress. Reset shared state only when the flag is False (top-level call). This ensures state is reset for standalone calls while preserving shared references for nested calls from hooks (which need access to the registry for cyclic structures).

Decoder (_decoding flag):

• decode(): set flag True, reset state, decode, set flag False
• decode_from_bytes(): reset state only when flag is False

Encoder (_encoding flag):

• encode(): set flag True, reset state, encode, set flag False
• encode_to_bytes(): reset state only when flag is False

Multiple vulnerabilities exist in cbor2 through version 5.7.0 in the decode_definite_long_string() function of the C extension decoder (source/decoder.c):

1. Integer Underflow Leading to Out-of-Bounds Read (CWE-191, CWE-125): An incorrect variable reference and missing state reset in the chunk processing loop causes buffer_length to not be reset to zero after UTF-8 character consumption. This results in subsequent chunk_length calculations producing negative values (e.g., chunk_length = 65536 - buffer_length), which are passed as signed integers to the read() method, potentially triggering unlimited read operations and resource exhaustion.
2. Memory Leak via Missing Reference Count Release (CWE-401): The main processing loop fails to release Python object references (Py_DECREF) for chunk objects allocated in each iteration. For CBOR strings longer than 65536 bytes, this causes cumulative memory leaks proportional to the payload size, enabling memory exhaustion attacks through repeated processing of large CBOR payloads.

Both vulnerabilities can be exploited remotely without authentication by sending specially-crafted CBOR data containing definite-length text strings with multi-byte UTF-8 characters positioned at 65536-byte chunk boundaries. Successful exploitation results in denial of service through process crashes (CBORDecodeEOF exceptions) or memory exhaustion. The vulnerabilities affect all applications using cbor2's C extension to process untrusted CBOR data, including web APIs, IoT data collectors, and message queue processors. Fixed in commit 851473490281f82d82560b2368284ef33cf6e8f9 pushed with released version 5.7.1.

Summary

• The cbor2 library is vulnerable to a Denial of Service (DoS) attack caused by uncontrolled recursion when decoding deeply nested CBOR structures.
• This vulnerability affects both the pure Python implementation and the C extension (_cbor2). The C extension correctly uses Python's C-API for recursion protection (Py_EnterRecursiveCall), but this mechanism is designed to prevent a stack overflow by raising a RecursionError. In some environments, this exception is not caught, thus causing the service process to terminate.
• While the library handles moderate nesting, it lacks a configurable, data-driven depth limit independent of Python's global recursion setting. An attacker can supply a crafted CBOR payload containing thousands of nested arrays (e.g., 0x81). When cbor2.loads() attempts to parse this, it hits the interpreter's recursion limit, causing the call to raise a RecursionError.
• By sending a stream of small (<100KB) malicious packets, an attacker can repeatedly crash worker processes faster than they can be restarted, resulting in a complete and sustained Denial of Service.

Details

• The vulnerability stems from the recursive design of the CBORDecoder class, specifically how it decodes nested container types like Arrays and Maps.
• Inside decode_array (and similarly decode_map), the decoder iterates through the number of elements specified in the CBOR header. For each element, it calls self.decode() again to parse the nested item. This recursive call lacks a depth-tracking mechanism.
• Vulnerable Code Locations:
  • cbor2/decoder.py (Pure Python implementation)
  • source/decoder.c (C extension implementation)
• Execution Flow:
  1. The cbor2.loads() function initializes a CBORDecoder and calls its decode() method.
  2. The decode() method reads the initial byte and dispatches control to a specific handler based on the major type. For an Array (Major Type 4), it calls decode_array.
  3. decode_array loops and calls self.decode() for each item, leading to deep recursion when parsing a payload like [...[...[1]...]...].

PoC

import cbor2

DEPTH = 1000

payload = b'\x81' * DEPTH + b'\x01'
print(f"[*] Payload size: {len(payload) / 1024:.2f} KB")
print("[*] Triggering decoder...")

try:
    cbor2.loads(payload)
    print("[+] Parsed successfully (Not Vulnerable)")
except RecursionError:
    print("\n[!] VULNERABLE: RecursionError triggered!")
except Exception as e:
    print(f"\n[-] Unexpected Error: {type(e).__name__}: {e}")

Impact

• Scope: This vulnerability affects any application using cbor2 to parse untrusted data. Common use cases include IoT data processing, WebAuthn (FIDO2) authentication flows, and inter-service communication over COSE (CBOR Object Signing and Encryption).
• Attack Vector: A remote, unauthenticated attacker can achieve a full Denial of Service with a highly efficient, low-bandwidth attack. A payload under 100KB is sufficient to reliably terminate a Python worker process.

Credit

This issue was discovered by Kevin Tu of TMIR at ByteDance. The patch was developed by @agronholm.

Summary

When a CBORDecoder instance is reused across multiple decode operations, values marked with the shareable tag (28) persist in memory and can be accessed by subsequent CBOR messages using the sharedref tag (29). This allows an attacker-controlled message to read data from previously decoded messages if the decoder is reused across trust boundaries.

Details

The issue is in the decoder's handling of the shareables list, which stores values tagged with CBOR tag 28 (shareable) for later reference by tag 29 (sharedref).

When decode_from_bytes() is called or when .fp is set to a new stream, the shareables list is not cleared. This allows references to persist across separate decode operations.

The issue exists in both the C extension and the pure Python decoder.

In the C extension (source/decoder.c), the _CBORDecoder_set_fp function (line ~202) updates the file pointer but does not reset the shareables state:

  static int
  _CBORDecoder_set_fp(CBORDecoderObject *self, PyObject *value, void *closure)
  {
      // ... validation ...
      tmp = self->read;
      self->read = read;
      Py_DECREF(tmp);
      return 0;
      // Missing: PyList_Clear(self->shareables) or equivalent
  }

In the pure Python decoder (cbor2/_decoder.py), the fp setter similarly fails to clear self._shareables.

Similarly, decode_from_bytes() in both implementations saves and restores the read pointer but does not clear the shareables list between decodes.

The shareable/sharedref tags are defined in the CBOR value sharing extension (http://cbor.schmorp.de/value-sharing) with scope limited to a single CBOR data item, not across separate messages.

PoC

import cbor2
from io import BytesIO

# Message from trusted source containing a shareable value
msg1 = cbor2.dumps(cbor2.CBORTag(28, "secret"))

# Attacker-controlled message referencing index 0
msg2 = cbor2.dumps(cbor2.CBORTag(29, 0))

# Decoder reused across trust boundaries
decoder = cbor2.CBORDecoder(BytesIO(b''))
decoder.decode_from_bytes(msg1)
print(decoder.decode_from_bytes(msg2))  # prints "secret"

No special configuration required. Affects any application that reuses a CBORDecoder instance to decode messages from different sources.

Impact

Information disclosure. Applications that reuse a CBORDecoder across trust boundaries are vulnerable if the trusted messages use value sharing (tag 28) and an attacker can send messages containing shared references (tag 29). An attacker who can send a crafted CBOR message containing a sharedref tag can read values from previously decoded messages, potentially exposing sensitive data such as credentials, tokens, or private user data.

Related

A similar issue in the encoder could produce invalid CBOR with dangling shared references:

import cbor2
from io import BytesIO

# Create encoder with value sharing enabled
encoder = cbor2.CBOREncoder(BytesIO(), value_sharing=True)

# Persistent object that will be encoded multiple times
shared_obj = ['hello']

# First encode: array containing shared_obj twice
encoder.encode([shared_obj, shared_obj])
print(f'First encode: {encoder.fp.getvalue().hex()}')
# Output: d81c82d81c816568656c6c6fd81d01

# Second encode: just shared_obj
encoder.fp = BytesIO()
encoder.encode(shared_obj)
result = encoder.fp.getvalue()
print(f'Second encode: {result.hex()}')
# Output: d81d01  (just a shared reference to index 1!)

# Try to decode the second result as standalone CBOR
decoder = cbor2.CBORDecoder(BytesIO(result))
decoded = decoder.decode()
# FAILS: shared reference 1 not found

While primarily a correctness bug, it could cause denial of service if invalid CBOR is transmitted to downstream systems that fail to parse it, or cause silent data corruption if the dangling reference happens to resolve to an unrelated value.

It can also be considered a memory leak in both the decoder and encoder as references are held that will never be released as long as the decoder/encoder remains alive.

Suggested resolution

Add dedicated boolean flags to track when an encode/decode operation is in progress. Reset shared state only when the flag is False (top-level call). This ensures state is reset for standalone calls while preserving shared references for nested calls from hooks (which need access to the registry for cyclic structures).

Decoder (_decoding flag):

• decode(): set flag True, reset state, decode, set flag False
• decode_from_bytes(): reset state only when flag is False

Encoder (_encoding flag):

• encode(): set flag True, reset state, encode, set flag False
• encode_to_bytes(): reset state only when flag is False

Multiple vulnerabilities exist in cbor2 through version 5.7.0 in the decode_definite_long_string() function of the C extension decoder (source/decoder.c):

1. Integer Underflow Leading to Out-of-Bounds Read (CWE-191, CWE-125): An incorrect variable reference and missing state reset in the chunk processing loop causes buffer_length to not be reset to zero after UTF-8 character consumption. This results in subsequent chunk_length calculations producing negative values (e.g., chunk_length = 65536 - buffer_length), which are passed as signed integers to the read() method, potentially triggering unlimited read operations and resource exhaustion.
2. Memory Leak via Missing Reference Count Release (CWE-401): The main processing loop fails to release Python object references (Py_DECREF) for chunk objects allocated in each iteration. For CBOR strings longer than 65536 bytes, this causes cumulative memory leaks proportional to the payload size, enabling memory exhaustion attacks through repeated processing of large CBOR payloads.

Both vulnerabilities can be exploited remotely without authentication by sending specially-crafted CBOR data containing definite-length text strings with multi-byte UTF-8 characters positioned at 65536-byte chunk boundaries. Successful exploitation results in denial of service through process crashes (CBORDecodeEOF exceptions) or memory exhaustion. The vulnerabilities affect all applications using cbor2's C extension to process untrusted CBOR data, including web APIs, IoT data collectors, and message queue processors. Fixed in commit 851473490281f82d82560b2368284ef33cf6e8f9 pushed with released version 5.7.1.

Summary

• The cbor2 library is vulnerable to a Denial of Service (DoS) attack caused by uncontrolled recursion when decoding deeply nested CBOR structures.
• This vulnerability affects both the pure Python implementation and the C extension (_cbor2). The C extension correctly uses Python's C-API for recursion protection (Py_EnterRecursiveCall), but this mechanism is designed to prevent a stack overflow by raising a RecursionError. In some environments, this exception is not caught, thus causing the service process to terminate.
• While the library handles moderate nesting, it lacks a configurable, data-driven depth limit independent of Python's global recursion setting. An attacker can supply a crafted CBOR payload containing thousands of nested arrays (e.g., 0x81). When cbor2.loads() attempts to parse this, it hits the interpreter's recursion limit, causing the call to raise a RecursionError.
• By sending a stream of small (<100KB) malicious packets, an attacker can repeatedly crash worker processes faster than they can be restarted, resulting in a complete and sustained Denial of Service.

Details

• The vulnerability stems from the recursive design of the CBORDecoder class, specifically how it decodes nested container types like Arrays and Maps.
• Inside decode_array (and similarly decode_map), the decoder iterates through the number of elements specified in the CBOR header. For each element, it calls self.decode() again to parse the nested item. This recursive call lacks a depth-tracking mechanism.
• Vulnerable Code Locations:
  • cbor2/decoder.py (Pure Python implementation)
  • source/decoder.c (C extension implementation)
• Execution Flow:
  1. The cbor2.loads() function initializes a CBORDecoder and calls its decode() method.
  2. The decode() method reads the initial byte and dispatches control to a specific handler based on the major type. For an Array (Major Type 4), it calls decode_array.
  3. decode_array loops and calls self.decode() for each item, leading to deep recursion when parsing a payload like [...[...[1]...]...].

PoC

import cbor2

DEPTH = 1000

payload = b'\x81' * DEPTH + b'\x01'
print(f"[*] Payload size: {len(payload) / 1024:.2f} KB")
print("[*] Triggering decoder...")

try:
    cbor2.loads(payload)
    print("[+] Parsed successfully (Not Vulnerable)")
except RecursionError:
    print("\n[!] VULNERABLE: RecursionError triggered!")
except Exception as e:
    print(f"\n[-] Unexpected Error: {type(e).__name__}: {e}")

Impact

• Scope: This vulnerability affects any application using cbor2 to parse untrusted data. Common use cases include IoT data processing, WebAuthn (FIDO2) authentication flows, and inter-service communication over COSE (CBOR Object Signing and Encryption).
• Attack Vector: A remote, unauthenticated attacker can achieve a full Denial of Service with a highly efficient, low-bandwidth attack. A payload under 100KB is sufficient to reliably terminate a Python worker process.

Credit

This issue was discovered by Kevin Tu of TMIR at ByteDance. The patch was developed by @agronholm.

Summary

When a CBORDecoder instance is reused across multiple decode operations, values marked with the shareable tag (28) persist in memory and can be accessed by subsequent CBOR messages using the sharedref tag (29). This allows an attacker-controlled message to read data from previously decoded messages if the decoder is reused across trust boundaries.

Details

The issue is in the decoder's handling of the shareables list, which stores values tagged with CBOR tag 28 (shareable) for later reference by tag 29 (sharedref).

When decode_from_bytes() is called or when .fp is set to a new stream, the shareables list is not cleared. This allows references to persist across separate decode operations.

The issue exists in both the C extension and the pure Python decoder.

In the C extension (source/decoder.c), the _CBORDecoder_set_fp function (line ~202) updates the file pointer but does not reset the shareables state:

  static int
  _CBORDecoder_set_fp(CBORDecoderObject *self, PyObject *value, void *closure)
  {
      // ... validation ...
      tmp = self->read;
      self->read = read;
      Py_DECREF(tmp);
      return 0;
      // Missing: PyList_Clear(self->shareables) or equivalent
  }

In the pure Python decoder (cbor2/_decoder.py), the fp setter similarly fails to clear self._shareables.

Similarly, decode_from_bytes() in both implementations saves and restores the read pointer but does not clear the shareables list between decodes.

The shareable/sharedref tags are defined in the CBOR value sharing extension (http://cbor.schmorp.de/value-sharing) with scope limited to a single CBOR data item, not across separate messages.

PoC

import cbor2
from io import BytesIO

# Message from trusted source containing a shareable value
msg1 = cbor2.dumps(cbor2.CBORTag(28, "secret"))

# Attacker-controlled message referencing index 0
msg2 = cbor2.dumps(cbor2.CBORTag(29, 0))

# Decoder reused across trust boundaries
decoder = cbor2.CBORDecoder(BytesIO(b''))
decoder.decode_from_bytes(msg1)
print(decoder.decode_from_bytes(msg2))  # prints "secret"

No special configuration required. Affects any application that reuses a CBORDecoder instance to decode messages from different sources.

Impact

Information disclosure. Applications that reuse a CBORDecoder across trust boundaries are vulnerable if the trusted messages use value sharing (tag 28) and an attacker can send messages containing shared references (tag 29). An attacker who can send a crafted CBOR message containing a sharedref tag can read values from previously decoded messages, potentially exposing sensitive data such as credentials, tokens, or private user data.

Related

A similar issue in the encoder could produce invalid CBOR with dangling shared references:

import cbor2
from io import BytesIO

# Create encoder with value sharing enabled
encoder = cbor2.CBOREncoder(BytesIO(), value_sharing=True)

# Persistent object that will be encoded multiple times
shared_obj = ['hello']

# First encode: array containing shared_obj twice
encoder.encode([shared_obj, shared_obj])
print(f'First encode: {encoder.fp.getvalue().hex()}')
# Output: d81c82d81c816568656c6c6fd81d01

# Second encode: just shared_obj
encoder.fp = BytesIO()
encoder.encode(shared_obj)
result = encoder.fp.getvalue()
print(f'Second encode: {result.hex()}')
# Output: d81d01  (just a shared reference to index 1!)

# Try to decode the second result as standalone CBOR
decoder = cbor2.CBORDecoder(BytesIO(result))
decoded = decoder.decode()
# FAILS: shared reference 1 not found

While primarily a correctness bug, it could cause denial of service if invalid CBOR is transmitted to downstream systems that fail to parse it, or cause silent data corruption if the dangling reference happens to resolve to an unrelated value.

It can also be considered a memory leak in both the decoder and encoder as references are held that will never be released as long as the decoder/encoder remains alive.

Suggested resolution

Add dedicated boolean flags to track when an encode/decode operation is in progress. Reset shared state only when the flag is False (top-level call). This ensures state is reset for standalone calls while preserving shared references for nested calls from hooks (which need access to the registry for cyclic structures).

Decoder (_decoding flag):

• decode(): set flag True, reset state, decode, set flag False
• decode_from_bytes(): reset state only when flag is False

Encoder (_encoding flag):

• encode(): set flag True, reset state, encode, set flag False
• encode_to_bytes(): reset state only when flag is False

Multiple vulnerabilities exist in cbor2 through version 5.7.0 in the decode_definite_long_string() function of the C extension decoder (source/decoder.c):

1. Integer Underflow Leading to Out-of-Bounds Read (CWE-191, CWE-125): An incorrect variable reference and missing state reset in the chunk processing loop causes buffer_length to not be reset to zero after UTF-8 character consumption. This results in subsequent chunk_length calculations producing negative values (e.g., chunk_length = 65536 - buffer_length), which are passed as signed integers to the read() method, potentially triggering unlimited read operations and resource exhaustion.
2. Memory Leak via Missing Reference Count Release (CWE-401): The main processing loop fails to release Python object references (Py_DECREF) for chunk objects allocated in each iteration. For CBOR strings longer than 65536 bytes, this causes cumulative memory leaks proportional to the payload size, enabling memory exhaustion attacks through repeated processing of large CBOR payloads.

Both vulnerabilities can be exploited remotely without authentication by sending specially-crafted CBOR data containing definite-length text strings with multi-byte UTF-8 characters positioned at 65536-byte chunk boundaries. Successful exploitation results in denial of service through process crashes (CBORDecodeEOF exceptions) or memory exhaustion. The vulnerabilities affect all applications using cbor2's C extension to process untrusted CBOR data, including web APIs, IoT data collectors, and message queue processors. Fixed in commit 851473490281f82d82560b2368284ef33cf6e8f9 pushed with released version 5.7.1.

Summary

• The cbor2 library is vulnerable to a Denial of Service (DoS) attack caused by uncontrolled recursion when decoding deeply nested CBOR structures.
• This vulnerability affects both the pure Python implementation and the C extension (_cbor2). The C extension correctly uses Python's C-API for recursion protection (Py_EnterRecursiveCall), but this mechanism is designed to prevent a stack overflow by raising a RecursionError. In some environments, this exception is not caught, thus causing the service process to terminate.
• While the library handles moderate nesting, it lacks a configurable, data-driven depth limit independent of Python's global recursion setting. An attacker can supply a crafted CBOR payload containing thousands of nested arrays (e.g., 0x81). When cbor2.loads() attempts to parse this, it hits the interpreter's recursion limit, causing the call to raise a RecursionError.
• By sending a stream of small (<100KB) malicious packets, an attacker can repeatedly crash worker processes faster than they can be restarted, resulting in a complete and sustained Denial of Service.

Details

• The vulnerability stems from the recursive design of the CBORDecoder class, specifically how it decodes nested container types like Arrays and Maps.
• Inside decode_array (and similarly decode_map), the decoder iterates through the number of elements specified in the CBOR header. For each element, it calls self.decode() again to parse the nested item. This recursive call lacks a depth-tracking mechanism.
• Vulnerable Code Locations:
  • cbor2/decoder.py (Pure Python implementation)
  • source/decoder.c (C extension implementation)
• Execution Flow:
  1. The cbor2.loads() function initializes a CBORDecoder and calls its decode() method.
  2. The decode() method reads the initial byte and dispatches control to a specific handler based on the major type. For an Array (Major Type 4), it calls decode_array.
  3. decode_array loops and calls self.decode() for each item, leading to deep recursion when parsing a payload like [...[...[1]...]...].

PoC

import cbor2

DEPTH = 1000

payload = b'\x81' * DEPTH + b'\x01'
print(f"[*] Payload size: {len(payload) / 1024:.2f} KB")
print("[*] Triggering decoder...")

try:
    cbor2.loads(payload)
    print("[+] Parsed successfully (Not Vulnerable)")
except RecursionError:
    print("\n[!] VULNERABLE: RecursionError triggered!")
except Exception as e:
    print(f"\n[-] Unexpected Error: {type(e).__name__}: {e}")

Impact

• Scope: This vulnerability affects any application using cbor2 to parse untrusted data. Common use cases include IoT data processing, WebAuthn (FIDO2) authentication flows, and inter-service communication over COSE (CBOR Object Signing and Encryption).
• Attack Vector: A remote, unauthenticated attacker can achieve a full Denial of Service with a highly efficient, low-bandwidth attack. A payload under 100KB is sufficient to reliably terminate a Python worker process.

Credit

This issue was discovered by Kevin Tu of TMIR at ByteDance. The patch was developed by @agronholm.

Summary

When a CBORDecoder instance is reused across multiple decode operations, values marked with the shareable tag (28) persist in memory and can be accessed by subsequent CBOR messages using the sharedref tag (29). This allows an attacker-controlled message to read data from previously decoded messages if the decoder is reused across trust boundaries.

Details

The issue is in the decoder's handling of the shareables list, which stores values tagged with CBOR tag 28 (shareable) for later reference by tag 29 (sharedref).

When decode_from_bytes() is called or when .fp is set to a new stream, the shareables list is not cleared. This allows references to persist across separate decode operations.

The issue exists in both the C extension and the pure Python decoder.

In the C extension (source/decoder.c), the _CBORDecoder_set_fp function (line ~202) updates the file pointer but does not reset the shareables state:

  static int
  _CBORDecoder_set_fp(CBORDecoderObject *self, PyObject *value, void *closure)
  {
      // ... validation ...
      tmp = self->read;
      self->read = read;
      Py_DECREF(tmp);
      return 0;
      // Missing: PyList_Clear(self->shareables) or equivalent
  }

In the pure Python decoder (cbor2/_decoder.py), the fp setter similarly fails to clear self._shareables.

Similarly, decode_from_bytes() in both implementations saves and restores the read pointer but does not clear the shareables list between decodes.

The shareable/sharedref tags are defined in the CBOR value sharing extension (http://cbor.schmorp.de/value-sharing) with scope limited to a single CBOR data item, not across separate messages.

PoC

import cbor2
from io import BytesIO

# Message from trusted source containing a shareable value
msg1 = cbor2.dumps(cbor2.CBORTag(28, "secret"))

# Attacker-controlled message referencing index 0
msg2 = cbor2.dumps(cbor2.CBORTag(29, 0))

# Decoder reused across trust boundaries
decoder = cbor2.CBORDecoder(BytesIO(b''))
decoder.decode_from_bytes(msg1)
print(decoder.decode_from_bytes(msg2))  # prints "secret"

No special configuration required. Affects any application that reuses a CBORDecoder instance to decode messages from different sources.

Impact

Information disclosure. Applications that reuse a CBORDecoder across trust boundaries are vulnerable if the trusted messages use value sharing (tag 28) and an attacker can send messages containing shared references (tag 29). An attacker who can send a crafted CBOR message containing a sharedref tag can read values from previously decoded messages, potentially exposing sensitive data such as credentials, tokens, or private user data.

Related

A similar issue in the encoder could produce invalid CBOR with dangling shared references:

import cbor2
from io import BytesIO

# Create encoder with value sharing enabled
encoder = cbor2.CBOREncoder(BytesIO(), value_sharing=True)

# Persistent object that will be encoded multiple times
shared_obj = ['hello']

# First encode: array containing shared_obj twice
encoder.encode([shared_obj, shared_obj])
print(f'First encode: {encoder.fp.getvalue().hex()}')
# Output: d81c82d81c816568656c6c6fd81d01

# Second encode: just shared_obj
encoder.fp = BytesIO()
encoder.encode(shared_obj)
result = encoder.fp.getvalue()
print(f'Second encode: {result.hex()}')
# Output: d81d01  (just a shared reference to index 1!)

# Try to decode the second result as standalone CBOR
decoder = cbor2.CBORDecoder(BytesIO(result))
decoded = decoder.decode()
# FAILS: shared reference 1 not found

While primarily a correctness bug, it could cause denial of service if invalid CBOR is transmitted to downstream systems that fail to parse it, or cause silent data corruption if the dangling reference happens to resolve to an unrelated value.

It can also be considered a memory leak in both the decoder and encoder as references are held that will never be released as long as the decoder/encoder remains alive.

Suggested resolution

Add dedicated boolean flags to track when an encode/decode operation is in progress. Reset shared state only when the flag is False (top-level call). This ensures state is reset for standalone calls while preserving shared references for nested calls from hooks (which need access to the registry for cyclic structures).

Decoder (_decoding flag):

• decode(): set flag True, reset state, decode, set flag False
• decode_from_bytes(): reset state only when flag is False

Encoder (_encoding flag):

• encode(): set flag True, reset state, encode, set flag False
• encode_to_bytes(): reset state only when flag is False

Multiple vulnerabilities exist in cbor2 through version 5.7.0 in the decode_definite_long_string() function of the C extension decoder (source/decoder.c):

1. Integer Underflow Leading to Out-of-Bounds Read (CWE-191, CWE-125): An incorrect variable reference and missing state reset in the chunk processing loop causes buffer_length to not be reset to zero after UTF-8 character consumption. This results in subsequent chunk_length calculations producing negative values (e.g., chunk_length = 65536 - buffer_length), which are passed as signed integers to the read() method, potentially triggering unlimited read operations and resource exhaustion.
2. Memory Leak via Missing Reference Count Release (CWE-401): The main processing loop fails to release Python object references (Py_DECREF) for chunk objects allocated in each iteration. For CBOR strings longer than 65536 bytes, this causes cumulative memory leaks proportional to the payload size, enabling memory exhaustion attacks through repeated processing of large CBOR payloads.

Both vulnerabilities can be exploited remotely without authentication by sending specially-crafted CBOR data containing definite-length text strings with multi-byte UTF-8 characters positioned at 65536-byte chunk boundaries. Successful exploitation results in denial of service through process crashes (CBORDecodeEOF exceptions) or memory exhaustion. The vulnerabilities affect all applications using cbor2's C extension to process untrusted CBOR data, including web APIs, IoT data collectors, and message queue processors. Fixed in commit 851473490281f82d82560b2368284ef33cf6e8f9 pushed with released version 5.7.1.

Summary

• The cbor2 library is vulnerable to a Denial of Service (DoS) attack caused by uncontrolled recursion when decoding deeply nested CBOR structures.
• This vulnerability affects both the pure Python implementation and the C extension (_cbor2). The C extension correctly uses Python's C-API for recursion protection (Py_EnterRecursiveCall), but this mechanism is designed to prevent a stack overflow by raising a RecursionError. In some environments, this exception is not caught, thus causing the service process to terminate.
• While the library handles moderate nesting, it lacks a configurable, data-driven depth limit independent of Python's global recursion setting. An attacker can supply a crafted CBOR payload containing thousands of nested arrays (e.g., 0x81). When cbor2.loads() attempts to parse this, it hits the interpreter's recursion limit, causing the call to raise a RecursionError.
• By sending a stream of small (<100KB) malicious packets, an attacker can repeatedly crash worker processes faster than they can be restarted, resulting in a complete and sustained Denial of Service.

Details

• The vulnerability stems from the recursive design of the CBORDecoder class, specifically how it decodes nested container types like Arrays and Maps.
• Inside decode_array (and similarly decode_map), the decoder iterates through the number of elements specified in the CBOR header. For each element, it calls self.decode() again to parse the nested item. This recursive call lacks a depth-tracking mechanism.
• Vulnerable Code Locations:
  • cbor2/decoder.py (Pure Python implementation)
  • source/decoder.c (C extension implementation)
• Execution Flow:
  1. The cbor2.loads() function initializes a CBORDecoder and calls its decode() method.
  2. The decode() method reads the initial byte and dispatches control to a specific handler based on the major type. For an Array (Major Type 4), it calls decode_array.
  3. decode_array loops and calls self.decode() for each item, leading to deep recursion when parsing a payload like [...[...[1]...]...].

PoC

import cbor2

DEPTH = 1000

payload = b'\x81' * DEPTH + b'\x01'
print(f"[*] Payload size: {len(payload) / 1024:.2f} KB")
print("[*] Triggering decoder...")

try:
    cbor2.loads(payload)
    print("[+] Parsed successfully (Not Vulnerable)")
except RecursionError:
    print("\n[!] VULNERABLE: RecursionError triggered!")
except Exception as e:
    print(f"\n[-] Unexpected Error: {type(e).__name__}: {e}")

Impact

• Scope: This vulnerability affects any application using cbor2 to parse untrusted data. Common use cases include IoT data processing, WebAuthn (FIDO2) authentication flows, and inter-service communication over COSE (CBOR Object Signing and Encryption).
• Attack Vector: A remote, unauthenticated attacker can achieve a full Denial of Service with a highly efficient, low-bandwidth attack. A payload under 100KB is sufficient to reliably terminate a Python worker process.

Credit

This issue was discovered by Kevin Tu of TMIR at ByteDance. The patch was developed by @agronholm.

Summary

When a CBORDecoder instance is reused across multiple decode operations, values marked with the shareable tag (28) persist in memory and can be accessed by subsequent CBOR messages using the sharedref tag (29). This allows an attacker-controlled message to read data from previously decoded messages if the decoder is reused across trust boundaries.

Details

The issue is in the decoder's handling of the shareables list, which stores values tagged with CBOR tag 28 (shareable) for later reference by tag 29 (sharedref).

When decode_from_bytes() is called or when .fp is set to a new stream, the shareables list is not cleared. This allows references to persist across separate decode operations.

The issue exists in both the C extension and the pure Python decoder.

In the C extension (source/decoder.c), the _CBORDecoder_set_fp function (line ~202) updates the file pointer but does not reset the shareables state:

  static int
  _CBORDecoder_set_fp(CBORDecoderObject *self, PyObject *value, void *closure)
  {
      // ... validation ...
      tmp = self->read;
      self->read = read;
      Py_DECREF(tmp);
      return 0;
      // Missing: PyList_Clear(self->shareables) or equivalent
  }

In the pure Python decoder (cbor2/_decoder.py), the fp setter similarly fails to clear self._shareables.

Similarly, decode_from_bytes() in both implementations saves and restores the read pointer but does not clear the shareables list between decodes.

The shareable/sharedref tags are defined in the CBOR value sharing extension (http://cbor.schmorp.de/value-sharing) with scope limited to a single CBOR data item, not across separate messages.

PoC

import cbor2
from io import BytesIO

# Message from trusted source containing a shareable value
msg1 = cbor2.dumps(cbor2.CBORTag(28, "secret"))

# Attacker-controlled message referencing index 0
msg2 = cbor2.dumps(cbor2.CBORTag(29, 0))

# Decoder reused across trust boundaries
decoder = cbor2.CBORDecoder(BytesIO(b''))
decoder.decode_from_bytes(msg1)
print(decoder.decode_from_bytes(msg2))  # prints "secret"

No special configuration required. Affects any application that reuses a CBORDecoder instance to decode messages from different sources.

Impact

Information disclosure. Applications that reuse a CBORDecoder across trust boundaries are vulnerable if the trusted messages use value sharing (tag 28) and an attacker can send messages containing shared references (tag 29). An attacker who can send a crafted CBOR message containing a sharedref tag can read values from previously decoded messages, potentially exposing sensitive data such as credentials, tokens, or private user data.

Related

A similar issue in the encoder could produce invalid CBOR with dangling shared references:

import cbor2
from io import BytesIO

# Create encoder with value sharing enabled
encoder = cbor2.CBOREncoder(BytesIO(), value_sharing=True)

# Persistent object that will be encoded multiple times
shared_obj = ['hello']

# First encode: array containing shared_obj twice
encoder.encode([shared_obj, shared_obj])
print(f'First encode: {encoder.fp.getvalue().hex()}')
# Output: d81c82d81c816568656c6c6fd81d01

# Second encode: just shared_obj
encoder.fp = BytesIO()
encoder.encode(shared_obj)
result = encoder.fp.getvalue()
print(f'Second encode: {result.hex()}')
# Output: d81d01  (just a shared reference to index 1!)

# Try to decode the second result as standalone CBOR
decoder = cbor2.CBORDecoder(BytesIO(result))
decoded = decoder.decode()
# FAILS: shared reference 1 not found

While primarily a correctness bug, it could cause denial of service if invalid CBOR is transmitted to downstream systems that fail to parse it, or cause silent data corruption if the dangling reference happens to resolve to an unrelated value.

It can also be considered a memory leak in both the decoder and encoder as references are held that will never be released as long as the decoder/encoder remains alive.

Suggested resolution

Add dedicated boolean flags to track when an encode/decode operation is in progress. Reset shared state only when the flag is False (top-level call). This ensures state is reset for standalone calls while preserving shared references for nested calls from hooks (which need access to the registry for cyclic structures).

Decoder (_decoding flag):

• decode(): set flag True, reset state, decode, set flag False
• decode_from_bytes(): reset state only when flag is False

Encoder (_encoding flag):

• encode(): set flag True, reset state, encode, set flag False
• encode_to_bytes(): reset state only when flag is False

Multiple vulnerabilities exist in cbor2 through version 5.7.0 in the decode_definite_long_string() function of the C extension decoder (source/decoder.c):

1. Integer Underflow Leading to Out-of-Bounds Read (CWE-191, CWE-125): An incorrect variable reference and missing state reset in the chunk processing loop causes buffer_length to not be reset to zero after UTF-8 character consumption. This results in subsequent chunk_length calculations producing negative values (e.g., chunk_length = 65536 - buffer_length), which are passed as signed integers to the read() method, potentially triggering unlimited read operations and resource exhaustion.
2. Memory Leak via Missing Reference Count Release (CWE-401): The main processing loop fails to release Python object references (Py_DECREF) for chunk objects allocated in each iteration. For CBOR strings longer than 65536 bytes, this causes cumulative memory leaks proportional to the payload size, enabling memory exhaustion attacks through repeated processing of large CBOR payloads.

Both vulnerabilities can be exploited remotely without authentication by sending specially-crafted CBOR data containing definite-length text strings with multi-byte UTF-8 characters positioned at 65536-byte chunk boundaries. Successful exploitation results in denial of service through process crashes (CBORDecodeEOF exceptions) or memory exhaustion. The vulnerabilities affect all applications using cbor2's C extension to process untrusted CBOR data, including web APIs, IoT data collectors, and message queue processors. Fixed in commit 851473490281f82d82560b2368284ef33cf6e8f9 pushed with released version 5.7.1.

Summary

• The cbor2 library is vulnerable to a Denial of Service (DoS) attack caused by uncontrolled recursion when decoding deeply nested CBOR structures.
• This vulnerability affects both the pure Python implementation and the C extension (_cbor2). The C extension correctly uses Python's C-API for recursion protection (Py_EnterRecursiveCall), but this mechanism is designed to prevent a stack overflow by raising a RecursionError. In some environments, this exception is not caught, thus causing the service process to terminate.
• While the library handles moderate nesting, it lacks a configurable, data-driven depth limit independent of Python's global recursion setting. An attacker can supply a crafted CBOR payload containing thousands of nested arrays (e.g., 0x81). When cbor2.loads() attempts to parse this, it hits the interpreter's recursion limit, causing the call to raise a RecursionError.
• By sending a stream of small (<100KB) malicious packets, an attacker can repeatedly crash worker processes faster than they can be restarted, resulting in a complete and sustained Denial of Service.

Details

• The vulnerability stems from the recursive design of the CBORDecoder class, specifically how it decodes nested container types like Arrays and Maps.
• Inside decode_array (and similarly decode_map), the decoder iterates through the number of elements specified in the CBOR header. For each element, it calls self.decode() again to parse the nested item. This recursive call lacks a depth-tracking mechanism.
• Vulnerable Code Locations:
  • cbor2/decoder.py (Pure Python implementation)
  • source/decoder.c (C extension implementation)
• Execution Flow:
  1. The cbor2.loads() function initializes a CBORDecoder and calls its decode() method.
  2. The decode() method reads the initial byte and dispatches control to a specific handler based on the major type. For an Array (Major Type 4), it calls decode_array.
  3. decode_array loops and calls self.decode() for each item, leading to deep recursion when parsing a payload like [...[...[1]...]...].

PoC

import cbor2

DEPTH = 1000

payload = b'\x81' * DEPTH + b'\x01'
print(f"[*] Payload size: {len(payload) / 1024:.2f} KB")
print("[*] Triggering decoder...")

try:
    cbor2.loads(payload)
    print("[+] Parsed successfully (Not Vulnerable)")
except RecursionError:
    print("\n[!] VULNERABLE: RecursionError triggered!")
except Exception as e:
    print(f"\n[-] Unexpected Error: {type(e).__name__}: {e}")

Impact

• Scope: This vulnerability affects any application using cbor2 to parse untrusted data. Common use cases include IoT data processing, WebAuthn (FIDO2) authentication flows, and inter-service communication over COSE (CBOR Object Signing and Encryption).
• Attack Vector: A remote, unauthenticated attacker can achieve a full Denial of Service with a highly efficient, low-bandwidth attack. A payload under 100KB is sufficient to reliably terminate a Python worker process.

Credit

This issue was discovered by Kevin Tu of TMIR at ByteDance. The patch was developed by @agronholm.

Summary

When a CBORDecoder instance is reused across multiple decode operations, values marked with the shareable tag (28) persist in memory and can be accessed by subsequent CBOR messages using the sharedref tag (29). This allows an attacker-controlled message to read data from previously decoded messages if the decoder is reused across trust boundaries.

Details

The issue is in the decoder's handling of the shareables list, which stores values tagged with CBOR tag 28 (shareable) for later reference by tag 29 (sharedref).

When decode_from_bytes() is called or when .fp is set to a new stream, the shareables list is not cleared. This allows references to persist across separate decode operations.

The issue exists in both the C extension and the pure Python decoder.

In the C extension (source/decoder.c), the _CBORDecoder_set_fp function (line ~202) updates the file pointer but does not reset the shareables state:

  static int
  _CBORDecoder_set_fp(CBORDecoderObject *self, PyObject *value, void *closure)
  {
      // ... validation ...
      tmp = self->read;
      self->read = read;
      Py_DECREF(tmp);
      return 0;
      // Missing: PyList_Clear(self->shareables) or equivalent
  }

In the pure Python decoder (cbor2/_decoder.py), the fp setter similarly fails to clear self._shareables.

Similarly, decode_from_bytes() in both implementations saves and restores the read pointer but does not clear the shareables list between decodes.

The shareable/sharedref tags are defined in the CBOR value sharing extension (http://cbor.schmorp.de/value-sharing) with scope limited to a single CBOR data item, not across separate messages.

PoC

import cbor2
from io import BytesIO

# Message from trusted source containing a shareable value
msg1 = cbor2.dumps(cbor2.CBORTag(28, "secret"))

# Attacker-controlled message referencing index 0
msg2 = cbor2.dumps(cbor2.CBORTag(29, 0))

# Decoder reused across trust boundaries
decoder = cbor2.CBORDecoder(BytesIO(b''))
decoder.decode_from_bytes(msg1)
print(decoder.decode_from_bytes(msg2))  # prints "secret"

No special configuration required. Affects any application that reuses a CBORDecoder instance to decode messages from different sources.

Impact

Information disclosure. Applications that reuse a CBORDecoder across trust boundaries are vulnerable if the trusted messages use value sharing (tag 28) and an attacker can send messages containing shared references (tag 29). An attacker who can send a crafted CBOR message containing a sharedref tag can read values from previously decoded messages, potentially exposing sensitive data such as credentials, tokens, or private user data.

Related

A similar issue in the encoder could produce invalid CBOR with dangling shared references:

import cbor2
from io import BytesIO

# Create encoder with value sharing enabled
encoder = cbor2.CBOREncoder(BytesIO(), value_sharing=True)

# Persistent object that will be encoded multiple times
shared_obj = ['hello']

# First encode: array containing shared_obj twice
encoder.encode([shared_obj, shared_obj])
print(f'First encode: {encoder.fp.getvalue().hex()}')
# Output: d81c82d81c816568656c6c6fd81d01

# Second encode: just shared_obj
encoder.fp = BytesIO()
encoder.encode(shared_obj)
result = encoder.fp.getvalue()
print(f'Second encode: {result.hex()}')
# Output: d81d01  (just a shared reference to index 1!)

# Try to decode the second result as standalone CBOR
decoder = cbor2.CBORDecoder(BytesIO(result))
decoded = decoder.decode()
# FAILS: shared reference 1 not found

While primarily a correctness bug, it could cause denial of service if invalid CBOR is transmitted to downstream systems that fail to parse it, or cause silent data corruption if the dangling reference happens to resolve to an unrelated value.

It can also be considered a memory leak in both the decoder and encoder as references are held that will never be released as long as the decoder/encoder remains alive.

Suggested resolution

Add dedicated boolean flags to track when an encode/decode operation is in progress. Reset shared state only when the flag is False (top-level call). This ensures state is reset for standalone calls while preserving shared references for nested calls from hooks (which need access to the registry for cyclic structures).

Decoder (_decoding flag):

• decode(): set flag True, reset state, decode, set flag False
• decode_from_bytes(): reset state only when flag is False

Encoder (_encoding flag):

• encode(): set flag True, reset state, encode, set flag False
• encode_to_bytes(): reset state only when flag is False

Multiple vulnerabilities exist in cbor2 through version 5.7.0 in the decode_definite_long_string() function of the C extension decoder (source/decoder.c):

1. Integer Underflow Leading to Out-of-Bounds Read (CWE-191, CWE-125): An incorrect variable reference and missing state reset in the chunk processing loop causes buffer_length to not be reset to zero after UTF-8 character consumption. This results in subsequent chunk_length calculations producing negative values (e.g., chunk_length = 65536 - buffer_length), which are passed as signed integers to the read() method, potentially triggering unlimited read operations and resource exhaustion.
2. Memory Leak via Missing Reference Count Release (CWE-401): The main processing loop fails to release Python object references (Py_DECREF) for chunk objects allocated in each iteration. For CBOR strings longer than 65536 bytes, this causes cumulative memory leaks proportional to the payload size, enabling memory exhaustion attacks through repeated processing of large CBOR payloads.

Both vulnerabilities can be exploited remotely without authentication by sending specially-crafted CBOR data containing definite-length text strings with multi-byte UTF-8 characters positioned at 65536-byte chunk boundaries. Successful exploitation results in denial of service through process crashes (CBORDecodeEOF exceptions) or memory exhaustion. The vulnerabilities affect all applications using cbor2's C extension to process untrusted CBOR data, including web APIs, IoT data collectors, and message queue processors. Fixed in commit 851473490281f82d82560b2368284ef33cf6e8f9 pushed with released version 5.7.1.

Summary

• The cbor2 library is vulnerable to a Denial of Service (DoS) attack caused by uncontrolled recursion when decoding deeply nested CBOR structures.
• This vulnerability affects both the pure Python implementation and the C extension (_cbor2). The C extension correctly uses Python's C-API for recursion protection (Py_EnterRecursiveCall), but this mechanism is designed to prevent a stack overflow by raising a RecursionError. In some environments, this exception is not caught, thus causing the service process to terminate.
• While the library handles moderate nesting, it lacks a configurable, data-driven depth limit independent of Python's global recursion setting. An attacker can supply a crafted CBOR payload containing thousands of nested arrays (e.g., 0x81). When cbor2.loads() attempts to parse this, it hits the interpreter's recursion limit, causing the call to raise a RecursionError.
• By sending a stream of small (<100KB) malicious packets, an attacker can repeatedly crash worker processes faster than they can be restarted, resulting in a complete and sustained Denial of Service.

Details

• The vulnerability stems from the recursive design of the CBORDecoder class, specifically how it decodes nested container types like Arrays and Maps.
• Inside decode_array (and similarly decode_map), the decoder iterates through the number of elements specified in the CBOR header. For each element, it calls self.decode() again to parse the nested item. This recursive call lacks a depth-tracking mechanism.
• Vulnerable Code Locations:
  • cbor2/decoder.py (Pure Python implementation)
  • source/decoder.c (C extension implementation)
• Execution Flow:
  1. The cbor2.loads() function initializes a CBORDecoder and calls its decode() method.
  2. The decode() method reads the initial byte and dispatches control to a specific handler based on the major type. For an Array (Major Type 4), it calls decode_array.
  3. decode_array loops and calls self.decode() for each item, leading to deep recursion when parsing a payload like [...[...[1]...]...].

PoC

import cbor2

DEPTH = 1000

payload = b'\x81' * DEPTH + b'\x01'
print(f"[*] Payload size: {len(payload) / 1024:.2f} KB")
print("[*] Triggering decoder...")

try:
    cbor2.loads(payload)
    print("[+] Parsed successfully (Not Vulnerable)")
except RecursionError:
    print("\n[!] VULNERABLE: RecursionError triggered!")
except Exception as e:
    print(f"\n[-] Unexpected Error: {type(e).__name__}: {e}")

Impact

• Scope: This vulnerability affects any application using cbor2 to parse untrusted data. Common use cases include IoT data processing, WebAuthn (FIDO2) authentication flows, and inter-service communication over COSE (CBOR Object Signing and Encryption).
• Attack Vector: A remote, unauthenticated attacker can achieve a full Denial of Service with a highly efficient, low-bandwidth attack. A payload under 100KB is sufficient to reliably terminate a Python worker process.

Credit

This issue was discovered by Kevin Tu of TMIR at ByteDance. The patch was developed by @agronholm.

Summary

When a CBORDecoder instance is reused across multiple decode operations, values marked with the shareable tag (28) persist in memory and can be accessed by subsequent CBOR messages using the sharedref tag (29). This allows an attacker-controlled message to read data from previously decoded messages if the decoder is reused across trust boundaries.

Details

The issue is in the decoder's handling of the shareables list, which stores values tagged with CBOR tag 28 (shareable) for later reference by tag 29 (sharedref).

When decode_from_bytes() is called or when .fp is set to a new stream, the shareables list is not cleared. This allows references to persist across separate decode operations.

The issue exists in both the C extension and the pure Python decoder.

In the C extension (source/decoder.c), the _CBORDecoder_set_fp function (line ~202) updates the file pointer but does not reset the shareables state:

  static int
  _CBORDecoder_set_fp(CBORDecoderObject *self, PyObject *value, void *closure)
  {
      // ... validation ...
      tmp = self->read;
      self->read = read;
      Py_DECREF(tmp);
      return 0;
      // Missing: PyList_Clear(self->shareables) or equivalent
  }

In the pure Python decoder (cbor2/_decoder.py), the fp setter similarly fails to clear self._shareables.

Similarly, decode_from_bytes() in both implementations saves and restores the read pointer but does not clear the shareables list between decodes.

The shareable/sharedref tags are defined in the CBOR value sharing extension (http://cbor.schmorp.de/value-sharing) with scope limited to a single CBOR data item, not across separate messages.

PoC

import cbor2
from io import BytesIO

# Message from trusted source containing a shareable value
msg1 = cbor2.dumps(cbor2.CBORTag(28, "secret"))

# Attacker-controlled message referencing index 0
msg2 = cbor2.dumps(cbor2.CBORTag(29, 0))

# Decoder reused across trust boundaries
decoder = cbor2.CBORDecoder(BytesIO(b''))
decoder.decode_from_bytes(msg1)
print(decoder.decode_from_bytes(msg2))  # prints "secret"

No special configuration required. Affects any application that reuses a CBORDecoder instance to decode messages from different sources.

Impact

Information disclosure. Applications that reuse a CBORDecoder across trust boundaries are vulnerable if the trusted messages use value sharing (tag 28) and an attacker can send messages containing shared references (tag 29). An attacker who can send a crafted CBOR message containing a sharedref tag can read values from previously decoded messages, potentially exposing sensitive data such as credentials, tokens, or private user data.

Related

A similar issue in the encoder could produce invalid CBOR with dangling shared references:

import cbor2
from io import BytesIO

# Create encoder with value sharing enabled
encoder = cbor2.CBOREncoder(BytesIO(), value_sharing=True)

# Persistent object that will be encoded multiple times
shared_obj = ['hello']

# First encode: array containing shared_obj twice
encoder.encode([shared_obj, shared_obj])
print(f'First encode: {encoder.fp.getvalue().hex()}')
# Output: d81c82d81c816568656c6c6fd81d01

# Second encode: just shared_obj
encoder.fp = BytesIO()
encoder.encode(shared_obj)
result = encoder.fp.getvalue()
print(f'Second encode: {result.hex()}')
# Output: d81d01  (just a shared reference to index 1!)

# Try to decode the second result as standalone CBOR
decoder = cbor2.CBORDecoder(BytesIO(result))
decoded = decoder.decode()
# FAILS: shared reference 1 not found

While primarily a correctness bug, it could cause denial of service if invalid CBOR is transmitted to downstream systems that fail to parse it, or cause silent data corruption if the dangling reference happens to resolve to an unrelated value.

It can also be considered a memory leak in both the decoder and encoder as references are held that will never be released as long as the decoder/encoder remains alive.

Suggested resolution

Add dedicated boolean flags to track when an encode/decode operation is in progress. Reset shared state only when the flag is False (top-level call). This ensures state is reset for standalone calls while preserving shared references for nested calls from hooks (which need access to the registry for cyclic structures).

Decoder (_decoding flag):

• decode(): set flag True, reset state, decode, set flag False
• decode_from_bytes(): reset state only when flag is False

Encoder (_encoding flag):

• encode(): set flag True, reset state, encode, set flag False
• encode_to_bytes(): reset state only when flag is False

Multiple vulnerabilities exist in cbor2 through version 5.7.0 in the decode_definite_long_string() function of the C extension decoder (source/decoder.c):

1. Integer Underflow Leading to Out-of-Bounds Read (CWE-191, CWE-125): An incorrect variable reference and missing state reset in the chunk processing loop causes buffer_length to not be reset to zero after UTF-8 character consumption. This results in subsequent chunk_length calculations producing negative values (e.g., chunk_length = 65536 - buffer_length), which are passed as signed integers to the read() method, potentially triggering unlimited read operations and resource exhaustion.
2. Memory Leak via Missing Reference Count Release (CWE-401): The main processing loop fails to release Python object references (Py_DECREF) for chunk objects allocated in each iteration. For CBOR strings longer than 65536 bytes, this causes cumulative memory leaks proportional to the payload size, enabling memory exhaustion attacks through repeated processing of large CBOR payloads.

Both vulnerabilities can be exploited remotely without authentication by sending specially-crafted CBOR data containing definite-length text strings with multi-byte UTF-8 characters positioned at 65536-byte chunk boundaries. Successful exploitation results in denial of service through process crashes (CBORDecodeEOF exceptions) or memory exhaustion. The vulnerabilities affect all applications using cbor2's C extension to process untrusted CBOR data, including web APIs, IoT data collectors, and message queue processors. Fixed in commit 851473490281f82d82560b2368284ef33cf6e8f9 pushed with released version 5.7.1.

Summary

• The cbor2 library is vulnerable to a Denial of Service (DoS) attack caused by uncontrolled recursion when decoding deeply nested CBOR structures.
• This vulnerability affects both the pure Python implementation and the C extension (_cbor2). The C extension correctly uses Python's C-API for recursion protection (Py_EnterRecursiveCall), but this mechanism is designed to prevent a stack overflow by raising a RecursionError. In some environments, this exception is not caught, thus causing the service process to terminate.
• While the library handles moderate nesting, it lacks a configurable, data-driven depth limit independent of Python's global recursion setting. An attacker can supply a crafted CBOR payload containing thousands of nested arrays (e.g., 0x81). When cbor2.loads() attempts to parse this, it hits the interpreter's recursion limit, causing the call to raise a RecursionError.
• By sending a stream of small (<100KB) malicious packets, an attacker can repeatedly crash worker processes faster than they can be restarted, resulting in a complete and sustained Denial of Service.

Details

• The vulnerability stems from the recursive design of the CBORDecoder class, specifically how it decodes nested container types like Arrays and Maps.
• Inside decode_array (and similarly decode_map), the decoder iterates through the number of elements specified in the CBOR header. For each element, it calls self.decode() again to parse the nested item. This recursive call lacks a depth-tracking mechanism.
• Vulnerable Code Locations:
  • cbor2/decoder.py (Pure Python implementation)
  • source/decoder.c (C extension implementation)
• Execution Flow:
  1. The cbor2.loads() function initializes a CBORDecoder and calls its decode() method.
  2. The decode() method reads the initial byte and dispatches control to a specific handler based on the major type. For an Array (Major Type 4), it calls decode_array.
  3. decode_array loops and calls self.decode() for each item, leading to deep recursion when parsing a payload like [...[...[1]...]...].

PoC

import cbor2

DEPTH = 1000

payload = b'\x81' * DEPTH + b'\x01'
print(f"[*] Payload size: {len(payload) / 1024:.2f} KB")
print("[*] Triggering decoder...")

try:
    cbor2.loads(payload)
    print("[+] Parsed successfully (Not Vulnerable)")
except RecursionError:
    print("\n[!] VULNERABLE: RecursionError triggered!")
except Exception as e:
    print(f"\n[-] Unexpected Error: {type(e).__name__}: {e}")

Impact

• Scope: This vulnerability affects any application using cbor2 to parse untrusted data. Common use cases include IoT data processing, WebAuthn (FIDO2) authentication flows, and inter-service communication over COSE (CBOR Object Signing and Encryption).
• Attack Vector: A remote, unauthenticated attacker can achieve a full Denial of Service with a highly efficient, low-bandwidth attack. A payload under 100KB is sufficient to reliably terminate a Python worker process.

Credit

This issue was discovered by Kevin Tu of TMIR at ByteDance. The patch was developed by @agronholm.

Summary

When a CBORDecoder instance is reused across multiple decode operations, values marked with the shareable tag (28) persist in memory and can be accessed by subsequent CBOR messages using the sharedref tag (29). This allows an attacker-controlled message to read data from previously decoded messages if the decoder is reused across trust boundaries.

Details

The issue is in the decoder's handling of the shareables list, which stores values tagged with CBOR tag 28 (shareable) for later reference by tag 29 (sharedref).

When decode_from_bytes() is called or when .fp is set to a new stream, the shareables list is not cleared. This allows references to persist across separate decode operations.

The issue exists in both the C extension and the pure Python decoder.

In the C extension (source/decoder.c), the _CBORDecoder_set_fp function (line ~202) updates the file pointer but does not reset the shareables state:

  static int
  _CBORDecoder_set_fp(CBORDecoderObject *self, PyObject *value, void *closure)
  {
      // ... validation ...
      tmp = self->read;
      self->read = read;
      Py_DECREF(tmp);
      return 0;
      // Missing: PyList_Clear(self->shareables) or equivalent
  }

In the pure Python decoder (cbor2/_decoder.py), the fp setter similarly fails to clear self._shareables.

Similarly, decode_from_bytes() in both implementations saves and restores the read pointer but does not clear the shareables list between decodes.

The shareable/sharedref tags are defined in the CBOR value sharing extension (http://cbor.schmorp.de/value-sharing) with scope limited to a single CBOR data item, not across separate messages.

PoC

import cbor2
from io import BytesIO

# Message from trusted source containing a shareable value
msg1 = cbor2.dumps(cbor2.CBORTag(28, "secret"))

# Attacker-controlled message referencing index 0
msg2 = cbor2.dumps(cbor2.CBORTag(29, 0))

# Decoder reused across trust boundaries
decoder = cbor2.CBORDecoder(BytesIO(b''))
decoder.decode_from_bytes(msg1)
print(decoder.decode_from_bytes(msg2))  # prints "secret"

No special configuration required. Affects any application that reuses a CBORDecoder instance to decode messages from different sources.

Impact

Information disclosure. Applications that reuse a CBORDecoder across trust boundaries are vulnerable if the trusted messages use value sharing (tag 28) and an attacker can send messages containing shared references (tag 29). An attacker who can send a crafted CBOR message containing a sharedref tag can read values from previously decoded messages, potentially exposing sensitive data such as credentials, tokens, or private user data.

Related

A similar issue in the encoder could produce invalid CBOR with dangling shared references:

import cbor2
from io import BytesIO

# Create encoder with value sharing enabled
encoder = cbor2.CBOREncoder(BytesIO(), value_sharing=True)

# Persistent object that will be encoded multiple times
shared_obj = ['hello']

# First encode: array containing shared_obj twice
encoder.encode([shared_obj, shared_obj])
print(f'First encode: {encoder.fp.getvalue().hex()}')
# Output: d81c82d81c816568656c6c6fd81d01

# Second encode: just shared_obj
encoder.fp = BytesIO()
encoder.encode(shared_obj)
result = encoder.fp.getvalue()
print(f'Second encode: {result.hex()}')
# Output: d81d01  (just a shared reference to index 1!)

# Try to decode the second result as standalone CBOR
decoder = cbor2.CBORDecoder(BytesIO(result))
decoded = decoder.decode()
# FAILS: shared reference 1 not found

While primarily a correctness bug, it could cause denial of service if invalid CBOR is transmitted to downstream systems that fail to parse it, or cause silent data corruption if the dangling reference happens to resolve to an unrelated value.

It can also be considered a memory leak in both the decoder and encoder as references are held that will never be released as long as the decoder/encoder remains alive.

Suggested resolution

Add dedicated boolean flags to track when an encode/decode operation is in progress. Reset shared state only when the flag is False (top-level call). This ensures state is reset for standalone calls while preserving shared references for nested calls from hooks (which need access to the registry for cyclic structures).

Decoder (_decoding flag):

• decode(): set flag True, reset state, decode, set flag False
• decode_from_bytes(): reset state only when flag is False

Encoder (_encoding flag):

• encode(): set flag True, reset state, encode, set flag False
• encode_to_bytes(): reset state only when flag is False

Multiple vulnerabilities exist in cbor2 through version 5.7.0 in the decode_definite_long_string() function of the C extension decoder (source/decoder.c):

1. Integer Underflow Leading to Out-of-Bounds Read (CWE-191, CWE-125): An incorrect variable reference and missing state reset in the chunk processing loop causes buffer_length to not be reset to zero after UTF-8 character consumption. This results in subsequent chunk_length calculations producing negative values (e.g., chunk_length = 65536 - buffer_length), which are passed as signed integers to the read() method, potentially triggering unlimited read operations and resource exhaustion.
2. Memory Leak via Missing Reference Count Release (CWE-401): The main processing loop fails to release Python object references (Py_DECREF) for chunk objects allocated in each iteration. For CBOR strings longer than 65536 bytes, this causes cumulative memory leaks proportional to the payload size, enabling memory exhaustion attacks through repeated processing of large CBOR payloads.

Both vulnerabilities can be exploited remotely without authentication by sending specially-crafted CBOR data containing definite-length text strings with multi-byte UTF-8 characters positioned at 65536-byte chunk boundaries. Successful exploitation results in denial of service through process crashes (CBORDecodeEOF exceptions) or memory exhaustion. The vulnerabilities affect all applications using cbor2's C extension to process untrusted CBOR data, including web APIs, IoT data collectors, and message queue processors. Fixed in commit 851473490281f82d82560b2368284ef33cf6e8f9 pushed with released version 5.7.1.

Summary

• The cbor2 library is vulnerable to a Denial of Service (DoS) attack caused by uncontrolled recursion when decoding deeply nested CBOR structures.
• This vulnerability affects both the pure Python implementation and the C extension (_cbor2). The C extension correctly uses Python's C-API for recursion protection (Py_EnterRecursiveCall), but this mechanism is designed to prevent a stack overflow by raising a RecursionError. In some environments, this exception is not caught, thus causing the service process to terminate.
• While the library handles moderate nesting, it lacks a configurable, data-driven depth limit independent of Python's global recursion setting. An attacker can supply a crafted CBOR payload containing thousands of nested arrays (e.g., 0x81). When cbor2.loads() attempts to parse this, it hits the interpreter's recursion limit, causing the call to raise a RecursionError.
• By sending a stream of small (<100KB) malicious packets, an attacker can repeatedly crash worker processes faster than they can be restarted, resulting in a complete and sustained Denial of Service.

Details

• The vulnerability stems from the recursive design of the CBORDecoder class, specifically how it decodes nested container types like Arrays and Maps.
• Inside decode_array (and similarly decode_map), the decoder iterates through the number of elements specified in the CBOR header. For each element, it calls self.decode() again to parse the nested item. This recursive call lacks a depth-tracking mechanism.
• Vulnerable Code Locations:
  • cbor2/decoder.py (Pure Python implementation)
  • source/decoder.c (C extension implementation)
• Execution Flow:
  1. The cbor2.loads() function initializes a CBORDecoder and calls its decode() method.
  2. The decode() method reads the initial byte and dispatches control to a specific handler based on the major type. For an Array (Major Type 4), it calls decode_array.
  3. decode_array loops and calls self.decode() for each item, leading to deep recursion when parsing a payload like [...[...[1]...]...].

PoC

import cbor2

DEPTH = 1000

payload = b'\x81' * DEPTH + b'\x01'
print(f"[*] Payload size: {len(payload) / 1024:.2f} KB")
print("[*] Triggering decoder...")

try:
    cbor2.loads(payload)
    print("[+] Parsed successfully (Not Vulnerable)")
except RecursionError:
    print("\n[!] VULNERABLE: RecursionError triggered!")
except Exception as e:
    print(f"\n[-] Unexpected Error: {type(e).__name__}: {e}")

Impact

• Scope: This vulnerability affects any application using cbor2 to parse untrusted data. Common use cases include IoT data processing, WebAuthn (FIDO2) authentication flows, and inter-service communication over COSE (CBOR Object Signing and Encryption).
• Attack Vector: A remote, unauthenticated attacker can achieve a full Denial of Service with a highly efficient, low-bandwidth attack. A payload under 100KB is sufficient to reliably terminate a Python worker process.

Credit

This issue was discovered by Kevin Tu of TMIR at ByteDance. The patch was developed by @agronholm.

Summary

When a CBORDecoder instance is reused across multiple decode operations, values marked with the shareable tag (28) persist in memory and can be accessed by subsequent CBOR messages using the sharedref tag (29). This allows an attacker-controlled message to read data from previously decoded messages if the decoder is reused across trust boundaries.

Details

The issue is in the decoder's handling of the shareables list, which stores values tagged with CBOR tag 28 (shareable) for later reference by tag 29 (sharedref).

When decode_from_bytes() is called or when .fp is set to a new stream, the shareables list is not cleared. This allows references to persist across separate decode operations.

The issue exists in both the C extension and the pure Python decoder.

In the C extension (source/decoder.c), the _CBORDecoder_set_fp function (line ~202) updates the file pointer but does not reset the shareables state:

  static int
  _CBORDecoder_set_fp(CBORDecoderObject *self, PyObject *value, void *closure)
  {
      // ... validation ...
      tmp = self->read;
      self->read = read;
      Py_DECREF(tmp);
      return 0;
      // Missing: PyList_Clear(self->shareables) or equivalent
  }

In the pure Python decoder (cbor2/_decoder.py), the fp setter similarly fails to clear self._shareables.

Similarly, decode_from_bytes() in both implementations saves and restores the read pointer but does not clear the shareables list between decodes.

The shareable/sharedref tags are defined in the CBOR value sharing extension (http://cbor.schmorp.de/value-sharing) with scope limited to a single CBOR data item, not across separate messages.

PoC

import cbor2
from io import BytesIO

# Message from trusted source containing a shareable value
msg1 = cbor2.dumps(cbor2.CBORTag(28, "secret"))

# Attacker-controlled message referencing index 0
msg2 = cbor2.dumps(cbor2.CBORTag(29, 0))

# Decoder reused across trust boundaries
decoder = cbor2.CBORDecoder(BytesIO(b''))
decoder.decode_from_bytes(msg1)
print(decoder.decode_from_bytes(msg2))  # prints "secret"

No special configuration required. Affects any application that reuses a CBORDecoder instance to decode messages from different sources.

Impact

Information disclosure. Applications that reuse a CBORDecoder across trust boundaries are vulnerable if the trusted messages use value sharing (tag 28) and an attacker can send messages containing shared references (tag 29). An attacker who can send a crafted CBOR message containing a sharedref tag can read values from previously decoded messages, potentially exposing sensitive data such as credentials, tokens, or private user data.

Related

A similar issue in the encoder could produce invalid CBOR with dangling shared references:

import cbor2
from io import BytesIO

# Create encoder with value sharing enabled
encoder = cbor2.CBOREncoder(BytesIO(), value_sharing=True)

# Persistent object that will be encoded multiple times
shared_obj = ['hello']

# First encode: array containing shared_obj twice
encoder.encode([shared_obj, shared_obj])
print(f'First encode: {encoder.fp.getvalue().hex()}')
# Output: d81c82d81c816568656c6c6fd81d01

# Second encode: just shared_obj
encoder.fp = BytesIO()
encoder.encode(shared_obj)
result = encoder.fp.getvalue()
print(f'Second encode: {result.hex()}')
# Output: d81d01  (just a shared reference to index 1!)

# Try to decode the second result as standalone CBOR
decoder = cbor2.CBORDecoder(BytesIO(result))
decoded = decoder.decode()
# FAILS: shared reference 1 not found

While primarily a correctness bug, it could cause denial of service if invalid CBOR is transmitted to downstream systems that fail to parse it, or cause silent data corruption if the dangling reference happens to resolve to an unrelated value.

It can also be considered a memory leak in both the decoder and encoder as references are held that will never be released as long as the decoder/encoder remains alive.

Suggested resolution

Add dedicated boolean flags to track when an encode/decode operation is in progress. Reset shared state only when the flag is False (top-level call). This ensures state is reset for standalone calls while preserving shared references for nested calls from hooks (which need access to the registry for cyclic structures).

Decoder (_decoding flag):

• decode(): set flag True, reset state, decode, set flag False
• decode_from_bytes(): reset state only when flag is False

Encoder (_encoding flag):

• encode(): set flag True, reset state, encode, set flag False
• encode_to_bytes(): reset state only when flag is False

Multiple vulnerabilities exist in cbor2 through version 5.7.0 in the decode_definite_long_string() function of the C extension decoder (source/decoder.c):

1. Integer Underflow Leading to Out-of-Bounds Read (CWE-191, CWE-125): An incorrect variable reference and missing state reset in the chunk processing loop causes buffer_length to not be reset to zero after UTF-8 character consumption. This results in subsequent chunk_length calculations producing negative values (e.g., chunk_length = 65536 - buffer_length), which are passed as signed integers to the read() method, potentially triggering unlimited read operations and resource exhaustion.
2. Memory Leak via Missing Reference Count Release (CWE-401): The main processing loop fails to release Python object references (Py_DECREF) for chunk objects allocated in each iteration. For CBOR strings longer than 65536 bytes, this causes cumulative memory leaks proportional to the payload size, enabling memory exhaustion attacks through repeated processing of large CBOR payloads.

Both vulnerabilities can be exploited remotely without authentication by sending specially-crafted CBOR data containing definite-length text strings with multi-byte UTF-8 characters positioned at 65536-byte chunk boundaries. Successful exploitation results in denial of service through process crashes (CBORDecodeEOF exceptions) or memory exhaustion. The vulnerabilities affect all applications using cbor2's C extension to process untrusted CBOR data, including web APIs, IoT data collectors, and message queue processors. Fixed in commit 851473490281f82d82560b2368284ef33cf6e8f9 pushed with released version 5.7.1.