AI Stack Alert High CVE-2026-116181 For AI Eng, SOC / IR, AppSec, CISO AI-assisted
vLLM 0.20.1–0.28.x: out-of-vocabulary stop_token_ids crash engine (CVE-2026-116181)
vLLM versions from 0.20.1 before 0.29.0 do not properly validate stop_token_ids. Remote attackers can send OpenAI chat, completion, generate, or batch requests with out-of-vocabulary IDs. This can terminate the engine or write outside the V2 runner's logits buffer. The CVSS base score is 7.1.
Why it matters
Any deployment exposing the vLLM OpenAI-compatible API to untrusted clients can have its inference engine crashed by an unauthenticated request. The out-of-bounds logits buffer write may also affect process memory integrity.
Do this
- Upgrade vLLM to 0.29.0 or later in all inference environments.
- Until patched, restrict access to the vLLM API endpoints to trusted clients at the gateway or network layer.
- Review request logs for chat, completion, generate, and batch calls that contain stop_token_ids values outside the model vocabulary range.
- Monitor for vLLM engine restarts or crashes and alert on repeated occurrences.
- Confirm the deployed version with the vLLM version command or container image tag after patching.
Exploit Clock
Exploitation unlikely (for now)
- EPSS 0.37%
Signals
- NoCISA KEV
- 7.1CVSS
- 0.37%EPSS
- NoPublic PoC
- 34/100Priority
- 70%Confidence
Evidence
“vLLM from 0.20.1 before 0.29.0 contains an improper input validation vulnerability that allows remote attackers to crash the engine”
“Attackers can send OpenAI chat, completion, generate, or batch requests with invalid IDs to terminate the engine or write outside the V2 runner's logits buffer.”
Sources
Affected / related
vLLM, stop_token_ids out-of-vocabulary input
Published 2026-10-25 14:16 UTC · updated 2026-11-08 02:48 UTC · AI-assisted summary, checked against quoted sources · How we score