vLLM before 0.29.0 Denial of Service via out-of-range stop_token_ids
cve-program · CVE-2026-100654 · Published
No access to your tenant or devices. Every item links to its source. What this means
Historical advisory recorded 23 Oct 2026 15:48:27 UTC. Read the latest advisory.
vLLM before 0.29.0 accepts user-controlled stop_token_ids on the OpenAI-compatible POST /v1/completions and POST /v1/chat/completions endpoints but validates only that the values are integers, not that each token id is within the model vocabulary/logits range. When min_tokens > 0, the stop token ids are used as logits indices to suppress stop tokens, so an out-of-range id reaches a CUDA indexing operation (index_put_) and triggers a device-side assertion. An authenticated API user can send a single malformed completion request that returns 500 Internal Server Error and puts EngineCore into a fatal state, causing subsequent requests to fail until the service is restarted (denial of service).
Read the publisher's advisorySee how the publisher's advisory changed over time. Each link opens the version saved at that time. The publisher's update date can differ from the date shown here.
Each link identifies one complete publisher record. History can include advisories that previously shared an identifier; an identifier removed by a publisher no longer establishes a current relationship. A historical link shows only that record and earlier saved versions. Failed checks and unchanged reads do not create new versions.
cve-program · CVE-2026-100654 · Published
This historical link shows only this version and versions saved before it. Browse all advisory history.