Executive brief
vLLM is a high-throughput engine for serving large language models (LLMs). A vulnerability in its API endpoints allows an unauthenticated remote attacker to crash the server or make it unresponsive by sending a single specially crafted web request. This could lead to service outages and disruption of AI-powered applications relying on the engine.
Technical details
A Denial of Service (DoS) vulnerability exists in vLLM due to uncontrolled resource consumption (CWE-400) when parsing HTTP headers. An unauthenticated remote attacker can send a single HTTP GET request containing an extremely large header (e.g., X-Forwarded-For) which the server attempts to load entirely into memory. This leads to system memory exhaustion and subsequent process crashes or unresponsiveness. The issue is fixed in version 0.10.1.1 by implementing appropriate HTTP limits. Users can also mitigate the risk by deploying a reverse proxy that enforces header size restrictions.
Affected products
- vLLM Project vllm >= 0.1.0, < 0.10.1.1
Timeline
- 2025-08-20: disclosed
- 2025-08-21: advisory
- 2025-08-21: patched: Version 0.10.1.1 released
References
- https://api.github.com/users/jperezdealgaba
- https://github.com/jperezdealgaba
- https://api.github.com/users/jperezdealgaba/gists%7B/gist_id%7D
- https://api.github.com/users/jperezdealgaba/repos
- https://avatars.githubusercontent.com/u/124347725?v=4
- https://api.github.com/users/jperezdealgaba/events%7B/privacy%7D