GHSA-8WJV-2P76-3863

Vulnerability from github – Published: 2026-09-29 23:43 – Updated: 2026-09-29 23:43
VLAI
Summary
PyJWT: Uncaught RecursionError in jwt.decode() on deeply nested token header
Details

Package

pyjwt (PyPI)

Affected versions

tested & verified on: 2.13.0. Every version whose PyJWS._load() translates only ValueError is affected

Description

PyJWS._looad() (jwt/api_jws.py:337) splits the compact token, base64url decodes the header segment and hands it to json.loads() before _verify_signature() runs. The parse is wrapped in except ValueError as e: raise DecodeError(...). That is what makes the documented contract sufficient for untrusted input: except jwt.PyJWTError (or InvalidTokenError) around jwt.decode(token, key, algorithms=[...]).

CPython's JSON decoder raises RecursionError (a RuntimeError subclass, not a ValueError) once nesting crosses the native stack guard. That exception leaves _load(), then jwt.decode(), and matches no PyJWTError handlerr at all. No key and no valid signature are needed: the header is parsed first, so a single unauthenticated request with an unsigned token suffices.

Scope: the recursion threshold is ~65,000 nesting levels; below it the same input yields a clean DecodeError (a 346,800-byte flat header is rejected normally, so depth, not size, is the trigger). A token this large cannot ride in an Authorization header (rejected at HTTP field limits,431 under gunicorn), but body transport (RFC 7662 introspection style) is a standard pattern. Under Flask 3.1.3 / gunicorn 26.2.0 each crafted request returns HTTP 500 instead of 401 and writes a full traceback to the error log; the sync worker survives.

Proof of concept

Against any endpoint that validates a token from its JSON body with the documented pattern:

import base64, json, http.client

def b64url(b): return base64.urlsafe_b64encode(b).rstrip(b"=")

# unsigned token: the header segment is 200,000 nested '[' (266,681 bytes total)
token = (b64url(b"[" * 200_000) + b"." + b64url(b'{"a":1}') + b"." + b64url(b"x")).decode()

conn = http.client.HTTPConnection("127.0.0.1", 8125, timeout=30)
conn.request("POST", "/api/protected", body=json.dumps({"token": token}),
             headers={"Content-Type": "application/json"})
print(conn.getresponse().status)   # 500, expected 401

Determinism: 20/20 crafted requests returned 500 against the reference container (PyJWT 2.13.0, Flask 3.1.3, gunicorn 26.2.0, Python 3.14.7) in 0.12 s (about 6 ms per request in a tight loop, 14 ms for a cold request), each leaving a RecursionError: Stack overflow ... while decoding a JSON array traceback in the error log.

Impact

An unauthenticated attacker can turn every request to a JWT validating endpoint into a server error with a full traceback logged per request, a cheap and repeatable denial of service on the authentication path (no crash, no data exposure).

Fix

Catch RecursionError alongside ValueError in _load() and raise DecodeError, or parse the header with a nonrecursive, depthbounded parser.

Reproduction

A standalone package (docker-compose.yml, Dockerfile, requirements.txt, target app, driver) accompanies this report:

cd <package dir>
docker compose up -d --build
python3 poc_check_then_attack.py
docker compose down -v

poc_pyjwt_recursion.zip

Exit 0= reproduced (control checks green, exactly one crafted request returns 500, traceback verified in docker logs, service alive afterwards); 1 = not reproduced; 2 = checks failed, no attack sent. The package's run_log.txt documents a complete proof run.

Credits

  • @Nivid42

Maintainer update — 2026-09-09

We reproduced the reported behavior on PyJWT 2.13.0: a deeply nested JSON array in the untrusted compact-JWS header reaches json.loads() before signature verification and raises RecursionError, which is not a PyJWTError. The same input is accepted by the public jwt.decode() path and can escape an application's normal except PyJWTError handling.

The fix is committed as 06573692ebcdec8831c3927513b3e87c31fbbb62. PyJWS._load() now translates both ValueError and RecursionError from header JSON parsing into DecodeError. A regression test covers the deeply nested-header path before signature verification. The full suite passes with 373 tests and 4 intentional cryptography-environment skips.

No release containing the fix has been published yet, so the patched version remains unset pending release planning. The advisory remains medium severity and unpublished.

Maintainer update — 2026-09-11

The verified fix for this advisory is included in PyJWT 2.14.0, released on 2026-09-11 and available on PyPI. PyJWT 2.14.0 is the first release containing the fix. This advisory is now published with 2.14.0 recorded as the patched version.

Show details on source website

{
  "affected": [
    {
      "package": {
        "ecosystem": "PyPI",
        "name": "pyJWT"
      },
      "ranges": [
        {
          "events": [
            {
              "introduced": "2.13.0"
            },
            {
              "fixed": "2.14.0"
            }
          ],
          "type": "ECOSYSTEM"
        }
      ]
    }
  ],
  "aliases": [
    "CVE-2026-102265"
  ],
  "database_specific": {
    "cwe_ids": [
      "CWE-674"
    ],
    "github_reviewed": true,
    "github_reviewed_at": "2026-09-29T23:43:06Z",
    "nvd_published_at": "2026-09-28T21:17:14Z",
    "severity": "MODERATE"
  },
  "details": "## Package\n\npyjwt (PyPI)\n\n## Affected versions\n\ntested \u0026 verified on: 2.13.0. Every version whose `PyJWS._load()` translates only `ValueError` is affected\n\n## Description\n\n`PyJWS._looad()` (`jwt/api_jws.py:337`) splits the compact token, base64url decodes the header segment and hands it to `json.loads()` before `_verify_signature()` runs. The parse is wrapped in `except ValueError as e: raise DecodeError(...)`. That is what makes the documented contract sufficient for untrusted input: `except jwt.PyJWTError` (or `InvalidTokenError`) around `jwt.decode(token, key, algorithms=[...])`.\n\nCPython\u0027s JSON decoder raises `RecursionError` (a `RuntimeError` subclass, not a `ValueError`) once nesting crosses the native stack guard. That exception leaves `_load()`, then `jwt.decode()`, and matches no `PyJWTError` handlerr at all. No key and no valid signature are needed: the header is parsed first, so a single unauthenticated request with an unsigned token suffices.\n\nScope: the recursion threshold is ~65,000 nesting levels; below it the same input yields a clean `DecodeError` (a 346,800-byte flat header is rejected normally, so depth, not size, is the trigger). A token this large cannot ride in an `Authorization` header (rejected at HTTP field limits,431 under gunicorn), but body transport (RFC 7662 introspection style) is a standard pattern. Under Flask 3.1.3 / gunicorn 26.2.0 each crafted request returns HTTP 500 instead of 401 and writes a full traceback to the error log; the sync worker survives.\n\n## Proof of concept\n\nAgainst any endpoint that validates a token from its JSON body with the documented pattern:\n\n```python\nimport base64, json, http.client\n\ndef b64url(b): return base64.urlsafe_b64encode(b).rstrip(b\"=\")\n\n# unsigned token: the header segment is 200,000 nested \u0027[\u0027 (266,681 bytes total)\ntoken = (b64url(b\"[\" * 200_000) + b\".\" + b64url(b\u0027{\"a\":1}\u0027) + b\".\" + b64url(b\"x\")).decode()\n\nconn = http.client.HTTPConnection(\"127.0.0.1\", 8125, timeout=30)\nconn.request(\"POST\", \"/api/protected\", body=json.dumps({\"token\": token}),\n             headers={\"Content-Type\": \"application/json\"})\nprint(conn.getresponse().status)   # 500, expected 401\n```\n\nDeterminism: 20/20 crafted requests returned 500 against the reference container (PyJWT 2.13.0, Flask 3.1.3, gunicorn 26.2.0, Python 3.14.7) in 0.12 s (about 6 ms per request in a tight loop, 14 ms for a cold request), each leaving a `RecursionError: Stack overflow ... while decoding a JSON array` traceback in the error log.\n\n## Impact\n\nAn unauthenticated attacker can turn every request to a JWT validating endpoint into a server error with a full traceback logged per request, a cheap and repeatable denial of service on the authentication path (no crash, no data exposure).\n\n## Fix\n\nCatch `RecursionError` alongside `ValueError` in `_load()` and raise `DecodeError`, or parse the header with a nonrecursive, depthbounded parser.\n\n## Reproduction\n\nA standalone package (docker-compose.yml, Dockerfile, requirements.txt, target app, driver) accompanies this report:\n\n```\ncd \u003cpackage dir\u003e\ndocker compose up -d --build\npython3 poc_check_then_attack.py\ndocker compose down -v\n```\n[poc_pyjwt_recursion.zip](https://github.com/user-attachments/files/31614987/poc_pyjwt_recursion.zip)\n\nExit 0= reproduced (control checks green, exactly one crafted request returns 500, traceback verified in `docker logs`, service alive afterwards); 1 = not reproduced; 2 = checks failed, no attack sent. The package\u0027s `run_log.txt` documents a complete proof run.\n\n## Credits\n\n- @Nivid42\n\n## Maintainer update \u2014 2026-09-09\n\nWe reproduced the reported behavior on PyJWT 2.13.0: a deeply nested JSON array in the untrusted compact-JWS header reaches `json.loads()` before signature verification and raises `RecursionError`, which is not a `PyJWTError`. The same input is accepted by the public `jwt.decode()` path and can escape an application\u0027s normal `except PyJWTError` handling.\n\nThe fix is committed as `06573692ebcdec8831c3927513b3e87c31fbbb62`. `PyJWS._load()` now translates both `ValueError` and `RecursionError` from header JSON parsing into `DecodeError`. A regression test covers the deeply nested-header path before signature verification. The full suite passes with 373 tests and 4 intentional cryptography-environment skips.\n\nNo release containing the fix has been published yet, so the patched version remains unset pending release planning. The advisory remains medium severity and unpublished.\n\n## Maintainer update \u2014 2026-09-11\n\nThe verified fix for this advisory is included in PyJWT 2.14.0, released on 2026-09-11 and available on PyPI. PyJWT 2.14.0 is the first release containing the fix. This advisory is now published with 2.14.0 recorded as the patched version.",
  "id": "GHSA-8wjv-2p76-3863",
  "modified": "2026-09-29T23:43:06Z",
  "published": "2026-09-29T23:43:06Z",
  "references": [
    {
      "type": "WEB",
      "url": "https://github.com/jpadilla/pyjwt/security/advisories/GHSA-8wjv-2p76-3863"
    },
    {
      "type": "ADVISORY",
      "url": "https://nvd.nist.gov/vuln/detail/CVE-2026-102265"
    },
    {
      "type": "WEB",
      "url": "https://github.com/jpadilla/pyjwt/commit/06573692ebcdec8831c3927513b3e87c31fbbb62"
    },
    {
      "type": "PACKAGE",
      "url": "https://github.com/jpadilla/pyjwt"
    },
    {
      "type": "WEB",
      "url": "https://github.com/jpadilla/pyjwt/releases/tag/2.14.0"
    }
  ],
  "schema_version": "1.4.0",
  "severity": [
    {
      "score": "CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:L",
      "type": "CVSS_V3"
    }
  ],
  "summary": "PyJWT: Uncaught RecursionError in jwt.decode() on deeply nested token header"
}



Log in or create an account to share your comment.




Tags
Taxonomy of the tags.


Loading…

Loading…

Loading…

Forecast uses a logistic model when the trend is rising, or an exponential decay model when the trend is falling. Fitted via linearized least squares.

Sightings

Author Source Type Date Other

Nomenclature

  • Seen: The vulnerability was mentioned, discussed, or observed by the user.
  • Confirmed: The vulnerability has been validated from an analyst's perspective.
  • Published Proof of Concept: A public proof of concept is available for this vulnerability.
  • Exploited: The vulnerability was observed as exploited by the user who reported the sighting.
  • Patched: The vulnerability was observed as successfully patched by the user who reported the sighting.
  • Not exploited: The vulnerability was not observed as exploited by the user who reported the sighting.
  • Not confirmed: The user expressed doubt about the validity of the vulnerability.
  • Not patched: The vulnerability was not observed as successfully patched by the user who reported the sighting.

Loading…

Loading…

Loading…

Related by attack behaviour

Vulnerabilities whose description is nearest to this one in the vector space of the CIRCL/vulnerability-attack-technique-biencoder model. This is a similarity search over the bi-encoder space (plain cosine), not a classification, and it has no measured accuracy.


Loading…