Tools AI Risk Radar ai-incident-0244
Incident record
Phi-4 audio input could crash a vLLM engine
- Severity
- Medium
- Status
- Patched
- Type
- Infra Vuln
- Target
- vllm-project/vllm
- Actor
- researcher
What happened
A long audio request could exhaust GPU memory in the Phi-4 multimodal encoder and crash the serving engine.
The disclosure concerns the eager-attention path and its retained attention matrices. The maintainer identifies vLLM 0.31.0 as affected and 0.32.0 as patched. The selected model and deployment’s authentication determine exposure.
Sources
- Maintainer security advisorygithub.com/vllm-project/vllm/security/advisories/GHSA-8397-8…
One record from the AI Risk Radar, maintained by the Security Desk.Data: CSV · JSON ·RSS · CC BY 4.0 with attribution to YFarmX.Tracker updated · 20 October 2026