PyJWT: HS256/384/512 verify accepts empty HMAC key via oct JWK (PyJWK/PyJWKClient) enabling token forgery

### Summary `HMACAlgorithm.from_jwk` accepts a symmetric (`kty=oct`) JWK whose `k` value is the empty string and returns a zero-length key without raising. When a token is then verified with that key object (a `PyJWK`, or a `PyJWK` resolved by `PyJWKSet` / `PyJWKClient`), the HMAC verifier computes `HMAC(key=b"", signing_input)` and accepts any token an attacker forged with the same empty key. The attacker needs no secret knowledge: they reproduce the digest locally with an empty key and set the token's `kid` to the empty-`k` JWK. The empty-key rejection that `HMACAlgorithm.prepare_key` performs for string/bytes secrets (added in 2.13.0) was never propagated to the sibling `from_jwk` code path, so the JWK/JWKS verification route stays exploitable. Under default decode options this path only emits an `InsecureKeyLengthWarning`, not an error, so verification still succeeds. This is the same "empty/nil HMAC key is accepted by verify" defect the JWT ecosystem is currently closing library-side: fast-jwt CVE-2026-44351 (CVSS 9.1) and ruby-jwt CVE-2026-45363 both added a library-side reject-empty-HMAC-key precondition. ### Affected versions `PyJWT` (PyPI package `pyjwt`) `<= 2.13.0`, current `master` at commit `7144e4534c34810f4525dc4578a32addd8212cff`. The string/bytes secret path is already guarded in 2.13.0; the JWK/JWKS path is not. ### Privilege required Unauthenticated. Any party who can reach a JWT-verifying endpoint of an application that verifies via a `PyJWK` / `PyJWKSet` / `PyJWKClient` whose oct key has an empty `k` can forge tokens with arbitrary claims. Preconditions: the verifier resolves the HMAC key from a JWK/JWKS whose oct entry carries an empty `k`. This occurs when the signing secret is sourced from an unset environment variable / unset config row / key-finder fallback that base64url-encodes `b""` into the published JWKS, or when the application trusts a JWKS containing such an entry. The equivalent unset-secret case for the string-secret API (`jwt.decode(token, key="")`) is already rejected in 2.13.0; the JWK route is the remaining reachable surface for the same misconfiguration. ### Vulnerable code `jwt/algorithms.py`, `HMACAlgorithm.from_jwk` (line 380): ```python @staticmethod def from_jwk(jwk: str | JWKDict) -> bytes: try: if isinstance(jwk, str): obj: JWKDict = json.loads(jwk) elif isinstance(jwk, dict): obj = jwk else: raise ValueError except ValueError: raise InvalidKeyError("Key is not valid JSON") from None if obj.get("kty") != "oct": raise InvalidKeyError("Not an HMAC key") return base64url_decode(obj["k"]) # <-- empty k -> b"", no rejection ``` Contrast the sibling `prepare_key`, which DID get the guard (lines 325-329): ```python def prepare_key(self, key: str | bytes) -> bytes: key_bytes = force_bytes(key) if len(key_bytes) == 0: raise InvalidKeyError("HMAC key must not be empty.") ``` In `jwt/api_jws.py`, `_verify_signature` (lines 389-416) uses the `PyJWK` key material directly, bypassing `prepare_key`, and the length warning is non-fatal by default: ```python if isinstance(key, PyJWK): if alg != key.algorithm_name: raise InvalidAlgorithmError( f"Token algorithm {alg!r} does not match the key's " f"algorithm {key.algorithm_name!r}" ) alg_obj = key.Algorithm prepared_key = key.key # from_jwk output (b""), prepare_key not run else: try: alg_obj = self.get_algorithm_by_name(alg) except NotImplementedError as e: raise InvalidAlgorithmError("Algorithm not supported") from e prepared_key = alg_obj.prepare_key(key) key_length_msg = alg_obj.check_key_length(prepared_key) if key_length_msg: if effective_options.get("enforce_minimum_key_length", False): raise InvalidKeyError(key_length_msg) else: warnings.warn(key_length_msg, InsecureKeyLengthWarning, stacklevel=4) # default: warn only if not alg_obj.verify(signing_input, prepared_key, signature): # verify with b"" succeeds raise InvalidSignatureError("Signature verification failed") ``` `PyJWK.__init__` sets `self.key = self.Algorithm.from_jwk(self._jwk_data)`, so a `{"kty":"oct","k":""}` JWK yields `PyJWK.key == b""`. `PyJWKSet` and `PyJWKClient.get_signing_key_from_jwt` build `PyJWK` objects the same way, so the JWKS-over-HTTP resolution route is affected identically. The 2.13.0 release note (`CHANGELOG.rst`) documents the `prepare_key` JWK/PEM hardening (GHSA-xgmm-8j9v-c9wx) and the `PyJWK` alg-binding fix (GHSA-jq35-7prp-9v3f), and the test suite has `test_hmac_prepare_key_rejects_empty_key` plus `test_hmac_from_jwk_should_raise_exception_if_empty_json` (which tests an empty JSON object `{}`, not an oct key with empty `k`). The empty-`k` oct case is the untested sibling path this report closes. ### How the attacker input reaches the sink An unauthenticated HTTP client controls the bearer token entirely: its `kid` selects the empty-`k` oct JWK, and its signature is computed locally with the empty key. The token flows request header -> `PyJWKClient.get_signing_key_from_jwt` (kid resolution against the trusted JWKS) -> `jwt.decode` with the resolved `PyJWK` -> `_verify_signature` -> `HMACAlgorithm.verify(signing_input, b"", sig)`. No secret, credential, or prior session is required; the only server-side precondition is a JWKS oct entry with an empty `k`, which the "unset signing secret" configuration produces. This is the identical trust boundary the string-secret path already guards. ### Proof of concept ### End-to-end reproduction against pinned PyPI `pyjwt` 2.13.0 (two deployed Flask services) Two services model a standard OIDC-style deployment: an auth service publishing a JWKS, and a pro…