Vulnerability  ·  2026-11-19

vLLM unauthenticated EngineCore crash via out-of-range prompt token id under --skip-tokenizer-init (CVE-2026-130685)

VulnerabilityMedium impactGlobalCVE-2026-130685
A single unauthenticated request with an out-of-range prompt token id crashes the shared EngineCore (and the API-server process) in vLLM when served with --skip-tokenizer-init. The vocab-range validation added for an earlier advisory (GHSA-25q3-v2hm-8vpf) lives in InputProcessor._validate_model_input but is gated on 'tokenizer is not None'; in this pre-tokenized/token-in-token-out/RL mode there is no tokenizer, so the check is skipped. The /v1/completions prompt field only constrains ids from below (ge=0), so a large positive id passes validation, reaches the embedding as an index, and raises out of the model forward; EngineCore treats that as fatal and shuts down the whole engine. Verified against vLLM main at HEAD dc89fdfb0e.
vLLM is a widely deployed open-source inference engine. Running with --skip-tokenizer-init (common for pre-tokenized / RL-serving workloads), one anonymous request permanently kills the shared engine for all tenants until manual restart — a trivial unauthenticated denial of service on the model-serving endpoint. The OpenAI-compatible server has no auth by default, so the whole multi-tenant serving layer is exposed.
Unauthenticated POST /v1/completions with a large out-of-range prompt token id → vocab check skipped → embedding index out-of-range → fatal exception in the shared EngineCore → permanent denial of service for all tenants.
vLLM 0.28.0 to before 0.33.0 (with --skip-tokenizer-init)
Upgrade to vLLM 0.33.0 or later (fix #62911 applies vocab-range check whenever prompt ids are present, not only when a tokenizer exists).
vLLM security advisory GHSA-q9v9-qpjp-fc24NVD detail — CVE-2026-130685
See this in the live feed Explore related AI security and governance findings — updated every morning.
Open the feed →