AI Stack Alert High CVE-2026-116178 For AI Eng, SOC / IR, AppSec, CISO AI-assisted
CVE-2026-116178: vLLM 0.20.0–0.30.0 remote DoS via malformed multimodal metadata
vLLM versions from 0.20.0 before 0.30.0 are affected by a denial of service flaw in the disaggregated /inference/v1/generate endpoint. Inconsistent multimodal metadata, such as repeated or overlapping mm_placeholders ranges or mismatched kwargs_data and mm_hashes lists, produces a negative text span in Qwen2.5-VL position construction. This terminates the engine. The CVSS base score is 8.2.
Why it matters
Remote, unauthenticated-level attackers who can reach the endpoint can crash the inference engine, taking down model serving for all users of that deployment. The source does not state the authentication requirement.
Do this
- Upgrade vLLM to 0.30.0 or later.
- Until patched, restrict network access to the disaggregated /inference/v1/generate endpoint to trusted callers.
- Monitor for vLLM engine crashes or restarts and correlate them with requests to /inference/v1/generate.
- Inspect logs for requests with repeated or overlapping mm_placeholders ranges or mismatched kwargs_data and mm_hashes lists.
Exploit Clock
Exploitation unlikely (for now)
- EPSS 0.33%
Signals
- NoCISA KEV
- 8.2CVSS
- 0.33%EPSS
- NoPublic PoC
- 34/100Priority
- 70%Confidence
Evidence
“vLLM from 0.20.0 before 0.30.0 contains a denial of service vulnerability that allows remote attackers to crash the engine”
“producing a negative text span in Qwen2.5-VL position construction that terminates the engine”
Sources
Affected / related
vLLM, Qwen2.5-VL, Denial of service via inconsistent multimodal metadata
Published 2026-10-25 14:16 UTC · updated 2026-11-08 02:49 UTC · AI-assisted summary, checked against quoted sources · How we score