CVE-2025-59425: vLLM timing attack in bearer authentication

Executive brief

vLLM, a library used for serving large language models, contains a security flaw in how it verifies API keys. An attacker could use a timing-based technique to guess a valid API key much faster than traditional guessing methods. If successful, this would allow unauthorized access to the AI models and data hosted on the server.

Technical details

A timing attack vulnerability (CWE-385) exists in vLLM's API server during bearer token validation. The root cause is the use of a standard string comparison for API key validation in `api_server.py`, which returns early upon finding a mismatch. This behavior allows a remote, unauthenticated attacker to measure response times to iteratively determine each character of a valid API key. Successful exploitation results in a full authentication bypass. The issue is resolved in version 0.11.0 by implementing constant-time string comparison.

Affected products

• vLLM Project vllm < 0.11.0

Timeline

• 2025-10-07: disclosed
• 2025-10-07: advisory
• 2025-10-07: patched: Fixed in version 0.11.0

References

• https://api.github.com/users/NiuBlibing
• https://github.com/NiuBlibing
• https://api.github.com/users/NiuBlibing/gists%7B/gist_id%7D
• https://api.github.com/users/NiuBlibing/repos
• https://avatars.githubusercontent.com/u/88433283?v=4
• https://api.github.com/users/NiuBlibing/events%7B/privacy%7D