vLLM versions prior to 0.27.0 contain a flaw in the handling of the DeepStream backend. The library fails to classify DeepStream as a GPU backend and removes pixel‑limit enforcement in the decode path. As a result, unauthenticated callers can trigger DeepStream during a request, which initializes a process‑wide GPU decode pool. The attacker can then submit video data that bypasses resource controls, producing a partial denial of service for other concurrent requests. The weakness is a classic resource exhaustion vulnerability, identified as CWE‑400.

The affected component is the vLLM library under the vllm‑project umbrella. All releases before 0.27.0 are vulnerable. Users of older versions should verify their local install and determine whether it includes the DeepStream path and supports GPU decoding.

The vulnerability is scored at CVSS 6.9, indicating a medium level of severity. The EPSS score is < 1%, and the vulnerability is not listed in the CISA KEV catalog, suggesting limited public exploitation data. The likely attack vector is an unauthenticated network request that activates DeepStream; a successful exploit would cause resource exhaustion and interrupt service availability for other users, but does not grant code execution or privileged access.