vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engine

Summary The Phi-4 multimodal audio encoder phi4mmaudio.py / phi4mmutils.py runs its conformer attention in eager mode by default: the official HF config of microsoft/Phi-4-multimodal-instruct does not set useptscaleddotproductattention, so the default False applies phi4mmaudio.py:166, :876;...

GHSA-8397-8QF6-F2GM vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engine

Summary The Phi-4 multimodal audio encoder phi4mmaudio.py / phi4mmutils.py runs its conformer attention in eager mode by default: the official HF config of microsoft/Phi-4-multimodal-instruct does not set useptscaleddotproductattention, so the default False applies phi4mmaudio.py:166, :876;...

CVE-2026-115913 vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engine

vLLM is an inference and serving engine for large language models. Prior to 0.32.0, the Phi-4 multimodal audio encoder in phi4mmaudio.py and phi4mmutils.py uses eager attention by default, allowing a remote client to submit one legal long audio clip that causes MultiHeadedAttention.forwardattenti...

CVE-2026-115913

Prior to version 0.32.0 , the vLLM inference and serving engine is vulnerable to a Denial of Service (DoS) condition via its Phi-4 multimodal audio encoder. Specifically, the components phi4mm_audio.py and phi4mm_utils.py use eager attention by default, causing MultiHeadedAttention.forward_attent...

CVE-2026-115913 vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engine

vLLM is an inference and serving engine for large language models. Prior to 0.32.0, the Phi-4 multimodal audio encoder in phi4mmaudio.py and phi4mmutils.py uses eager attention by default, allowing a remote client to submit one legal long audio clip that causes MultiHeadedAttention.forwardattenti...

PT-2026-123436

Name of the Vulnerable Software and Affected Versions vLLM versions prior to 0.32.0 Description The Phi-4 multimodal audio encoder in phi4mm audio.py and phi4mm utils.py uses eager attention by default. This allows a remote client to submit a long audio clip that triggers the...

Research on Models Engaging in Genie-Like Behavior

New paper: "Self-Jailbreaking: Language Models Can Reason Themselves Out of Safety Alignment After Benign Reasoning Training." Abstract: We discover a novel and surprising phenomenon of unintentional misalignment in reasoning language models RLMs, which we call self-jailbreaking. Specifically,...