AI Stack Alert Medium CVE-2026-115916 For AI Eng, SOC / IR, AppSec, CISO AI-assisted
vLLM < 0.32.0: out-of-vocabulary prompt token id crashes engine (CVE-2026-115916)
vLLM versions before 0.32.0 accept a pre-tokenized prompt token id equal to the vocabulary size for models whose vocab is a multiple of 64, such as Llama-2 (vocab 32000). At tensor-parallel size 1 the id reaches an embedding lookup out of range, raising a CUDA device-side assert that takes the engine down. The fix is in 0.32.0.
Why it matters
Any client that can send pre-tokenized input can crash the inference engine with one request, a denial of service. Deployments running tp=1 with affected models are exposed.
Do this
- Upgrade vLLM to 0.32.0 or later.
- Until patched, reject or filter token-id prompts (for example /v1/completions with prompt as token ids) that contain ids >= the model vocabulary size.
- Identify deployments running tensor-parallel size 1 with Llama-2-family or other vocab-size-multiple-of-64 models and prioritize them for upgrade.
- Monitor engine logs for CUDA device-side assert errors and engine restarts as indicators of this issue.
Exploit Clock
Exploitation unlikely (for now)
- EPSS 0.31%
Signals
- NoCISA KEV
- 6.5CVSS
- 0.31%EPSS
- 27/100Priority
- 85%Confidence
Evidence
“a single request carrying a pre-tokenized prompt with token id equal to the vocab size (32000) passes vLLM's input-id validation”
“which poisons the engine's CUDA context and takes the engine down”
“Fixed in [MR59491](https://github.com/vllm-project/vllm/pull/59491), released in v0.32.0.”
Sources
Affected / related
vLLM, Llama-2 family, out-of-vocabulary token id input validation bypass
Published 2026-10-28 21:15 UTC · updated 2026-11-08 02:42 UTC · AI-assisted summary, checked against quoted sources · How we score