The vulnerability is an input validation flaw in the cache_salt parameter accepted by the OpenAI‑compatible and Anthropic‑compatible endpoints of vLLM. The parameter accepts arbitrarily large strings and is processed on a single EngineCore scheduler thread. When an unauthenticated attacker sends a request with a multi‑hundred‑megabyte cache_salt value, the system performs expensive pickle serialization and SHA‑256 hashing, causing the scheduler thread to stall and denying service to all concurrent requests.

vLLM packages before version 0.29.0, any deployment of the vllm project that exposes the OpenAI‑compatible or Anthropic‑compatible API endpoints, regardless of hosting environment. The flaw resides in the core library and is triggered by external HTTP requests.

The CVSS score of 6.9 indicates a medium severity denial‑of‑service risk. The flaw is accessed through an API parameter, so any unauthenticated user can exploit it by sending arbitrary HTTP requests. The EPSS score is not available, and the vulnerability is not listed in CISA's KEV catalog, suggesting no known public exploits yet. Attackers could achieve full service disruption by flooding the API with large cache_salt values, forcing the scheduler thread to consume CPU resources while it serializes and hashes the payload.