Information Exposure Through Timing Discrepancy in vLLM - CVE-2025-46570

Detailed vulnerability description

The vulnerability allows a remote user to disclose sensitive information.

The vulnerability exists due to a timing side-channel in the chunk-based prefix caching mechanism when processing prompts that share cached prefix chunks. A remote user can measure time to first token for crafted prompt guesses to disclose sensitive information.

Exploitation requires sharing the same backend with a victim, and user interaction is required.

How to mitigate CVE-2025-46570

Sources

• https://github.com/vllm-project/vllm/security/advisories/GHSA-4qjh-9fv9-r85r
• https://github.com/vllm-project/vllm/pull/17045