AI Stack Alert Medium CVE-2026-115913 For AI Eng, SOC / IR, AppSec AI-assisted
vLLM <0.32.0: one max-length audio request crashes engine via Phi-4 multimodal encoder (CVE-2026-115913)
The Phi-4 multimodal audio encoder in vLLM runs eager attention by default and stores the attention matrix on the module, so the matrices of all completed layers stay alive. A single 10-minute audio clip, which is within the default limits, can exhaust GPU memory and crash the engine. The issue is fixed in vLLM 0.32.0.
Why it matters
Any unauthenticated client that can reach the chat or completions API of a default deployment serving Phi-4-multimodal-instruct can take the whole inference server down with one request. The result is a full denial of service that requires a process restart.
Do this
- Upgrade vLLM to 0.32.0 or later on all hosts serving Phi-4-multimodal-instruct or other models that reuse this audio encoder.
- Until patched, lower VLLM_MAX_AUDIO_DECODE_DURATION_S and VLLM_MAX_AUDIO_CLIP_FILESIZE_MB below the defaults (600 s and 25 MB) to reduce the largest accepted audio input.
- Limit or disable audio input on public endpoints for affected models at the API gateway or proxy layer.
- Alert on EngineDeadError, CUDA out-of-memory errors, and unexpected vLLM process restarts, and correlate them with audio-bearing requests.
Exploit Clock
Exploitation unlikely (for now)
- EPSS 0.33%
Signals
- NoCISA KEV
- 6.5CVSS
- 0.33%EPSS
- 26/100Priority
- 75%Confidence
Evidence
“The eager path fully materializes the attention probability matrix of shape”
“carrying a perfectly legal 10-minute audio clip (silent noise suffices — no crafted content needed).”
Sources
Affected / related
vLLM, Phi-4-multimodal-instruct
Published 2026-10-28 21:15 UTC · updated 2026-11-08 02:43 UTC · AI-assisted summary, checked against quoted sources · How we score