6.5
MediumCVSS 3.1 · from GitHub_M
CVE-2026-115916
vLLM: Out-of-vocabulary prompt token id passes input validation and triggers a CUDA device-side assert (engine crash / DoS)
vllm · CWE-129 · CWE-129Affectedvllm
FixNo fix listed yet.
PublishedOct 26, 17:18 · GitHub_M
ScoredOct 26 · 6.5 · GitHub_M
Exploit code publicNone known · checked 21:18 UTC
ExploitedNone known · CISA KEV checked 22:11 UTC
What we know
No known exploitation. CISA KEV, checked 22:11 UTC.
No public exploit code known. Exploit-DB, Metasploit, Nuclei, checked 21:18 UTC.
An attacker needs network access and an account. CVSS: network, low privileges.
A successful attack can take the service down. CVSS impact.
Description
vLLM is an inference and serving engine for large language models. Prior to 0.32.0, vllm/tokenizers/hf.py computes max_token_id with an off-by-one upper bound, allowing a pre-tokenized prompt ID equal to the vocabulary size to pass validation. For a tensor-parallel size of one and an exact multiple-of-64 vocabulary, VocabParallelEmbedding.forward does not mask that value before F.embedding uses it, causing a CUDA device-side assertion and terminating the engine. This issue is fixed in version 0.32.0.
From GitHub_M
Affected and fixed versions
ProductAffectedFixed invllm< 0.32.0–
Exploit code
Exploit-DBnone knownMetasploitnone knownNucleinone known