AI Stack Alert High CVE-2026-126397 For CISO, SOC / IR, AppSec, AI Eng AI-assisted
vLLM 0.10.1 to 0.31.0: request-controlled max_patches enables CPU denial of service on Cohere2 Vision
vLLM versions from 0.10.1 before 0.31.0 pass a request-controlled mm_processor_kwargs.max_patches value to the Hugging Face processor for Cohere2 Vision models. The source reports no server-side upper bound on this value. Remote clients can send small image requests with large, varying max_patches values to bypass the LRU cache and exhaust preprocessing workers.
Why it matters
Any remote client that can reach an affected vLLM endpoint can degrade service for other users. The source assigns a CVSS base score of 8.7.
Do this
- Upgrade vLLM to version 0.31.0 or later.
- Until patched, restrict access to vLLM chat-completion endpoints to trusted clients.
- Reject or cap mm_processor_kwargs.max_patches at the API gateway or proxy layer for Cohere2 Vision models.
- Monitor preprocessing worker CPU use and request latency for sustained spikes tied to image requests.
Exploit Clock
No signals yet
- No exploitation signals yet
Signals
- NoCISA KEV
- 8.7CVSS
- NoPublic PoC
- 36/100Priority
- 80%Confidence
Evidence
“vLLM from 0.10.1 before 0.31.0 forwards a request-controlled mm_processor_kwargs.max_patches value to the Hugging Face processor for Cohere2 Vision models without a server-side upper bound”
“Remote clients can send small image chat-completion requests with large, varying max_patches values to bypass the LRU cache, exhaust preprocessing workers, and deny service to other clients.”
Sources
Affected / related
vLLM, Cohere2 Vision
Published 2026-11-13 10:17 UTC · updated 2026-11-13 12:35 UTC · AI-assisted summary, checked against quoted sources · How we score