CVE-2026-115913: vLLM out-of-memory crash via long audio clip to Phi-4 multimodal encoder

Summary

vLLM versions prior to 0.32.0 use eager attention by default in the Phi-4 multimodal audio encoder (phi4mm_audio.py and phi4mm_utils.py). A remote client can submit a single legal long audio clip, causing MultiHeadedAttention.forward_attention to materialize and retain full attention matrices in self.attn. The resulting quadratic GPU memory use can trigger an out-of-memory condition that terminates EngineCore for all users.

Mitigation

Fixed in version 0.32.0.

Topics

• Inference infrastructure

Related items

• MediumCVE-2026-115911: vLLM chat template rendering denial of service via unbounded Jinja executionSimilar attack · NVD/CVE Database
• MediumCVE-2026-116181: vLLM improper input validation of stop_token_ids crashes engineSimilar attack · NVD/CVE Database
• HighCVE-2026-116180: vLLM denial of service through prompt_logprobs on cached prefixesSimilar attack · NVD/CVE Database
• HighCVE-2026-116179: vLLM resource exhaustion through media_io_kwargs video frame-count overrideSimilar attack · NVD/CVE Database
• MediumCVE-2026-116178: vLLM denial of service via inconsistent multimodal metadataSimilar attack · NVD/CVE Database