Low severity3.1NVD Advisory· Published Sep 19, 2026· Updated Sep 28, 2026
CVE-2026-93989
CVE-2026-93989
Description
vLLM through 0.29.0 fails to properly validate bad_words token indices against the model's generation output width in SamplingParams.update_from_tokenizer(). Attackers can supply out-of-bounds token indices that corrupt logits memory of concurrent requests, causing different in-flight HTTP requests to return incorrect tokens.
AI Insight
LLM-synthesized narrative grounded in this CVE's description and references.
Affected products
3Patches
Vulnerability mechanics
References
4- github.com/vllm-project/vllm/pull/48824nvdIssue TrackingPatch
- www.vulncheck.com/advisories/vllm-through-0.29.0-cross-request-logits-corruption-via-bad-wordsnvdThird Party Advisory
- github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/sampling_params.pynvdProduct
- github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/v1/worker/gpu/sample/bad_words.pynvdProduct
News mentions
1- VLLM: Batch of 17 Vulnerabilities Disclosed, Affecting LLM Inference Engine StabilityVypr Intelligence · Sep 21, 2026