AI Stack Alert High CVE-2026-126399 For CISO, SOC / IR, AppSec, AI Eng AI-assisted
CVE-2026-126399: vLLM 0.19.0 to before 0.31.0 resource exhaustion via video num_frames bypass
vLLM from 0.19.0 before 0.31.0 contains a resource exhaustion vulnerability. Inference clients can bypass the server frame limit by setting media_io_kwargs.video.num_frames in chat completion requests. This forces VideoMediaIO to decode all frames and stack them, causing memory exhaustion.
Why it matters
Any exposed vLLM inference endpoint accepting chat completion requests with video input can be driven to memory exhaustion, causing denial of service for other users of the server. The CVSS base score is 7.1.
Do this
- Upgrade vLLM to version 0.31.0 or later.
- Until patched, reject or cap media_io_kwargs.video.num_frames values at the gateway, including negative values and values of -1.
- Limit request body size and the number of inline base64 media items per chat completion request at the reverse proxy.
- Monitor vLLM worker memory use and alert on sudden growth during video chat completion requests.
Exploit Clock
No signals yet
- No exploitation signals yet
Signals
- NoCISA KEV
- 7.1CVSS
- NoPublic PoC
- 36/100Priority
- 80%Confidence
Evidence
“vLLM from 0.19.0 before 0.31.0 contains a resource exhaustion vulnerability that allows inference clients to bypass the server frame limit”
“forcing VideoMediaIO to decode all frames and np.stack them, causing memory exhaustion.”
Sources
Affected / related
vLLM, resource exhaustion via frame limit bypass
Published 2026-11-13 10:17 UTC · updated 2026-11-13 12:35 UTC · AI-assisted summary, checked against quoted sources · How we score