GPU VulnDB

Database/AI/ML frameworks & serving

vLLM: out-of-range prompt token ids crash the shared EngineCore when tokenizer init is skipped

CVSS 8.7CVE-2026-130685AI/ML frameworks & servingcurated

Impact

When vLLM runs with --skip-tokenizer-init, the vocabulary range check on prompt token ids is skipped. A single request carrying a large positive token id to /v1/completions, /v1/embeddings or /pooling causes a fatal embedding index error in the shared EngineCore. Every client served by that engine loses service, and the GPUs behind it sit idle until the server is restarted. Only deployments that use --skip-tokenizer-init are affected.

Who can reach it

Anyone who can send requests to the vLLM API endpoint. No authentication is needed unless the operator puts it in front.

What to do

Upgrade vLLM to 0.33.0 or later and restart the serving process. Until then, drop --skip-tokenizer-init where you can, or validate token ids against the model vocabulary in a gateway in front of the server.

References

Related entries

All AI/ML frameworks & serving entries

This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.