{"id":"CVE-2026-115913","kind":"vulnerability","aliases":["CVE-2026-115913"],"related":[],"title":"vLLM: one long Phi-4 audio clip can exhaust GPU memory and crash","description":"vLLM is an inference and serving engine for large language models. Prior to 0.32.0, the Phi-4 multimodal audio encoder in phi4mm_audio.py and phi4mm_utils.py uses eager attention by default, allowing a remote client to submit one legal long audio clip that causes MultiHeadedAttention.forward_attention to materialize and retain full attention matrices in self.attn. The scaled dot-product attention branch does not materialize these matrices. The resulting quadratic GPU-memory use can cause an out-of-memory condition and terminate EngineCore for all users. This issue is fixed in version 0.32.0.","summary":"vLLM's Phi-4 multimodal audio encoder uses eager attention by default, so one long audio clip from a remote client makes it store full quadratic attention matrices in GPU memory. The out-of-memory condition terminates EngineCore and interrupts service for all users.","severity":"medium","cvss":6.5,"epss":null,"exploited":false,"exploited_at":null,"products":[],"ecosystems":[],"tags":["ai-tooling","gpu-inference","phi-4","python","self-hosted","vllm"],"not_affected_if":["no-remote-audio-input","text-only-models"],"how_to_check":"- Check the installed version with `pip show vllm` and treat anything below `0.32.0` as affected.\n- Check whether the deployment serves the Phi-4 multimodal model with audio input enabled.","fix":"- Upgrade vLLM to `0.32.0` or later.","status":"open","published_at":"2026-10-26T17:17:19+00:00","first_seen_at":"2026-10-26T17:31:22+00:00","triaged_at":"2026-10-26T17:33:17+00:00","updated_at":"2026-10-26T17:33:17+00:00","category":"ai-tooling","confidence":0.7,"triage":{"at":"2026-10-26T17:33:17Z","model":"claude-haiku-5-5","output":{"fix":"- Upgrade vLLM to `0.32.0` or later.","reach":"internet","title":"vLLM: one long Phi-4 audio clip can exhaust GPU memory and crash","impact":"medium","summary":"vLLM's Phi-4 multimodal audio encoder uses eager attention by default, so one long audio clip from a remote client makes it store full quadratic attention matrices in GPU memory. The out-of-memory condition terminates EngineCore and interrupts service for all users.","category":"ai-tooling","nicknames":[],"who_cares":["vllm","self-hosted","gpu-inference","phi-4","python"],"confidence":0.7,"needs_auth":"unknown","worst_case":"A single request crashes the inference engine and denies service to all users.","analyst_note":"","how_to_check":"- Check the installed version with `pip show vllm` and treat anything below `0.32.0` as affected.\n- Check whether the deployment serves the Phi-4 multimodal model with audio input enabled.","needs_analyst":false,"not_affected_if":["text-only-models","no-remote-audio-input"],"who_is_affected":"Teams self-hosting vLLM below 0.32.0 that serve the Phi-4 multimodal model and accept audio input from clients they do not fully trust. Deployments that serve only text models, or that do not accept audio from remote clients, are not exposed to this path.","user_interaction":"no","needs_special_setup":"yes","severity_if_unknown":"medium"},"cost_usd":0.000598,"input_tokens":3407,"output_tokens":515,"severity_before":"medium"},"analyst_status":null,"analyst_note":null,"search":"'-115913':16A,19A '-2026':15A,18A '-4':5A,27B,36B,94C '0.32.0':91C,170C 'ai':21B 'ai-tool':20B 'allow':107C 'attent':42B,59B,104C,121C,127C,136C 'audio':6A,38B,48B,96C,116C 'branch':137C 'caus':119C,151C 'client':53B,110C 'clip':7A,49B,117C 'condit':69B,157C 'crash':13A 'cve':14A,17A 'default':44B,106C 'dot':134C 'dot-product':133C 'eager':41B,103C 'encod':39B,97C 'engin':84C 'enginecor':71B,160C 'exhaust':9A 'fix':167C 'full':57B,126C 'gpu':10A,24B,62B,147C 'gpu-infer':23B 'gpu-memori':146C 'host':31B 'infer':25B,81C 'interrupt':73B 'issu':165C 'languag':87C 'larg':86C 'legal':114C 'long':3A,47B,115C 'make':54B 'materi':123C,140C 'matric':60B,128C,142C 'memori':11A,63B,68B,148C,156C 'model':88C 'multiheadedattention.forward':120C 'multimod':37B,95C 'one':2A,46B,113C 'out-of-memori':65B,153C 'phi':4A,26B,35B,93C 'phi4mm_audio.py':99C 'phi4mm_utils.py':101C 'prior':89C 'product':135C 'python':28B 'quadrat':58B,145C 'remot':52B,109C 'result':144C 'retain':125C 'scale':132C 'self':30B 'self-host':29B 'self.attn':130C 'serv':83C 'servic':74B 'store':56B 'submit':112C 'termin':70B,159C 'tool':22B 'use':40B,102C,149C 'user':77B,163C 'version':169C 'vllm':1A,32B,33B,78C","nicknames":[],"mitigations":[],"who_affected":"Teams self-hosting vLLM below 0.32.0 that serve the Phi-4 multimodal model and accept audio input from clients they do not fully trust. Deployments that serve only text models, or that do not accept audio from remote clients, are not exposed to this path.","tier":"B","verified":{},"analyst_revisit_at":null,"first_public_at":"2026-10-26T17:17:19+00:00","kev":null,"cvss_vector":"CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H","cvss_by":"security-advisories@github.com","poc_count":0,"sources":[{"source":"nvd","external_id":"CVE-2026-115913","url":"https://nvd.nist.gov/vuln/detail/CVE-2026-115913","severity":null,"published_at":"2026-10-26T17:17:19+00:00","modified_at":"2026-10-26T17:17:19+00:00","first_seen_at":"2026-10-26T17:31:22+00:00","fetched_at":"2026-10-26T17:31:22+00:00","note":null}],"documents":[],"notice":"Free to read at a human pace. An API with keys and an MCP server are in preparation; early access: hello@infrastack.ai"}