Denial of Service Vulnerability in vLLM by VLLM Project

CVE-2025-29770

What is CVE-2025-29770?

The vLLM high-throughput inference engine features a cache for compiled grammars utilized by its outlines library. This cache, enabled by default, is susceptible to abuse by malicious users who can exploit the system by sending numerous short decoding requests with unique schemas. Each request adds to the cache, potentially leading to a Denial of Service if the filesystem capacity is exceeded. This issue arises despite vLLM's default configuration, which may not use the outlines backend, as it can still be selected on a request basis. A fix is implemented in version 0.8.0.

References