CVE-2025-59425: vLLM: timing attack enables API key bypass

vLLM is an inference and serving engine for large language models (LLMs). Before version 0.11.0rc2, the API key support in vLLM performs validation using a method that was vulnerable to a timing attack. API key validation uses a string comparison that takes longer the more characters the provided API key gets correct. Data analysis across many attempts could allow an attacker to determine when it finds the next correct character in the key sequence. Deployments relying on vLLM's built-in API key validation are vulnerable to authentication bypass using this technique. Version 0.11.0rc2 fixes the issue.

An adversary discovers a vLLM instance via passive internet scanning (Shodan, Censys — vLLM's OpenAI-compatible /health endpoint is a clear fingerprint). They write a Python script that iterates over each character position of the API key, sending batches of requests with candidate characters and measuring median response time across each batch. Because Python's == operator compares strings character-by-character and exits early on mismatch, correct characters produce measurably longer comparison times. After O(N * charset_size * samples) requests — fully automatable in hours on a fast connection — the full API key is reconstructed. The attacker then authenticates legitimately and queries the model for sensitive system prompt extraction, runs cost-harvesting inference loops, or uses the access as initial foothold for lateral movement.