CVE-2026-93841
LOWCVSS 3.7Description
vLLM through 0.29.0 contains a memory corruption vulnerability in the Triton _bincount_kernel where prompt token IDs index the penalty prompt-presence bitset without bounds checking against vocabulary size. Attackers can submit multimodal audio requests with tokens equal to vocabulary size, causing out-of-bounds writes that corrupt concurrent requests' sampler state and alter repetition penalty behavior.
Timeline
Published:September 18th, 2026 8:17 PM
Last modified:September 28th, 2026 6:35 PM
CVSS Scoring
CVSS v3: 3.7 (LOW)
CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:N/I:L/A:N
Affected Software
From NVD CPE configuration data
Vendors
vllm
Products
vllm
Weaknesses (CWE)
References & Reports
Advisories, patches, and third-party reports
- https://github.com/vllm-project/vllmProduct
- https://github.com/vllm-project/vllm/blob/v0.29.0/vllm/v1/worker/gpu/sample/penalties.py#L241-L255Product
- https://github.com/vllm-project/vllm/blob/v0.29.0/vllm/v1/worker/gpu/sample/penalties.py#L33-L37Product
- https://github.com/vllm-project/vllm/pull/49081Issue Tracking
- https://www.vulncheck.com/advisories/vllm-through-0.29.0-adjacent-request-sampler-state-corruption-via-unvalidated-prompt-token-idsThird Party Advisory
Source: NIST NVD · Data may lag official sources by up to one minute