{"items":[{"id":"NVI-00008994","kind":"vulnerability","title":"vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engine","revision":3,"unknowns":["Customer inventory, deployment, reachability and exploitation have not been inspected."],"lifecycle":"active","identifiers":["GHSA-8397-8qf6-f2gm","CVE-2026-115913"],"plain_summary":"GitHub lists this reviewed advisory for vllm. Published version ranges and source prerequisites are shown below; this is not an inspection of a visitor’s application.","review_status":"source_listed","control_topics":[],"public_actions":[{"summary":"Identify the exact component and deployed version, examine the linked publisher’s conditions, and retain evidence of any mitigation.","source_ids":["sr-353794dc676f151430d20f747410a389"],"action_type":"investigate","limitations":"Confirm current vendor support, compatibility and deployment evidence. This report has not tested a fix in your system.","current_supported_target":null}],"schema_version":"1.1.0","source_metrics":[{"kind":"CVSS","as_of":"2026-10-28T21:15:42.000Z","value":"6.5 medium","meaning":"GitHub source severity; exploitability and local consequence require separate evidence.","source_id":"sr-65e85a5ff56321ef9af773e89e4d758d"},{"kind":"EPSS","as_of":"2026-11-01T12:00:16.000Z","value":"0.329%","meaning":"FIRST EPSS model v2026.06.15; exploitation probability over 30 days, percentile 0.2611. Missing scores are unknown.","source_id":"sr-353794dc676f151430d20f747410a389"}],"why_it_matters":"Published affected context helps an owner decide what evidence to gather. This report has not examined your application.","related_sources":[{"id":"sr-65e85a5ff56321ef9af773e89e4d758d","url":"https://github.com/advisories/GHSA-8397-8qf6-f2gm","basis":"vendor_advisory","notes":"GitHub Advisory Database; CC-BY-4.0. Each advisory links its original attribution. Null normalizes fields.","title":"GHSA-8397-8qf6-f2gm","publisher":"GitHub Advisory Database","observed_at":"2026-10-31T21:56:08.420Z"},{"id":"sr-353794dc676f151430d20f747410a389","url":"https://www.first.org/epss/data.html","basis":"official_catalog","notes":"FIRST EPSS; probability of exploitation in the next 30 days, not local compromise. Dataset date/model retained; FIRST services terms apply.","title":"CVE-2026-115913","publisher":"FIRST EPSS","observed_at":"2026-11-01T14:28:32.267Z"}],"affected_context":[{"certainty":"source_listed","source_ids":["sr-65e85a5ff56321ef9af773e89e4d758d"],"technology":"vllm","known_exclusions":[],"conditions_summary":"### Summary\n\nThe Phi-4 multimodal audio encoder (`phi4mm_audio.py` / `phi4mm_utils.py`) runs its conformer attention in **eager mode by default**: the official HF config of `microsoft/Phi-4-multimodal-instruct` does not set `use_pt_scaled_dot_product_attention`, so the default `False` applies (`phi4mm_audio.py:166`, `:876`; `phi4mm_utils.py:1669`). The eager path fully materializes the attention probability matrix of shape `[n_chunks, 16, 500, 500]` per layer — and then **stores it on the module**: `self.attn = attn` (`phi4mm_utils.py:1802`, a `torch.jit.Attribute` declared at `:1693`), keeping a live reference for every one of the 24 encoder layers.\n\nThe only gates on audio length are the server-side limits `VLLM_MAX_AUDIO_CLIP_FILESIZE_MB=25` and `VLLM_MAX_AUDIO_DECODE_DURATION_S=600` (`envs.py:83-84`) — i.e. a 10-minute clip is a **legal input by default**. A 600 s clip becomes ~60,000 mel frames, which the encoder unfolds into 500-frame chunks (`phi4mm_audio.py:996`, `max_seq_len = 500`), giving `n_chunks ≈ 120`. Each layer therefore materializes a `120×16×500×500` matrix (~1–2 GiB depending on dtype), and the module attribute keeps all completed layers' matrices alive, so GPU ","public_affected_ranges":["< 0.32.0"]}],"null_published_at":"2026-11-07T18:43:15.098Z","source_updated_at":"2026-10-28T21:15:42.000Z","source_disclosed_at":"2026-10-28T21:15:41.000Z","is_synthetic_fixture":false,"null_last_reviewed_at":null,"public_change_summary":"Dated EPSS estimate updated; this is not a new disclosure or proof of exploitation.","null_first_observed_at":"2026-10-31T21:56:08.420Z"},{"id":"NVI-00008994","kind":"vulnerability","title":"vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engine","revision":2,"unknowns":["Customer inventory, deployment, reachability and exploitation have not been inspected."],"lifecycle":"active","identifiers":["GHSA-8397-8qf6-f2gm","CVE-2026-115913"],"plain_summary":"GitHub lists this reviewed advisory for vllm. Published version ranges and source prerequisites are shown below; this is not an inspection of a visitor’s application.","review_status":"source_listed","control_topics":[],"public_actions":[{"summary":"Identify the exact component and deployed version, examine the linked publisher’s conditions, and retain evidence of any mitigation.","source_ids":["sr-030336df5db5af49982b2e108c55e6cb"],"action_type":"investigate","limitations":"Confirm current vendor support, compatibility and deployment evidence. This report has not tested a fix in your system.","current_supported_target":null}],"schema_version":"1.1.0","source_metrics":[{"kind":"CVSS","as_of":"2026-10-28T21:15:42.000Z","value":"6.5 medium","meaning":"GitHub source severity; exploitability and local consequence require separate evidence.","source_id":"sr-65e85a5ff56321ef9af773e89e4d758d"},{"kind":"EPSS","as_of":"2026-10-31T12:00:27.000Z","value":"0.329%","meaning":"FIRST EPSS model v2026.06.15; exploitation probability over 30 days, percentile 0.26117. Missing scores are unknown.","source_id":"sr-030336df5db5af49982b2e108c55e6cb"}],"why_it_matters":"Published affected context helps an owner decide what evidence to gather. This report has not examined your application.","related_sources":[{"id":"sr-65e85a5ff56321ef9af773e89e4d758d","url":"https://github.com/advisories/GHSA-8397-8qf6-f2gm","basis":"vendor_advisory","notes":"GitHub Advisory Database; CC-BY-4.0. Each advisory links its original attribution. Null normalizes fields.","title":"GHSA-8397-8qf6-f2gm","publisher":"GitHub Advisory Database","observed_at":"2026-10-31T21:56:08.420Z"},{"id":"sr-030336df5db5af49982b2e108c55e6cb","url":"https://www.first.org/epss/data.html","basis":"official_catalog","notes":"FIRST EPSS; probability of exploitation in the next 30 days, not local compromise. Dataset date/model retained; FIRST services terms apply.","title":"CVE-2026-115913","publisher":"FIRST EPSS","observed_at":"2026-10-31T22:06:15.374Z"}],"affected_context":[{"certainty":"source_listed","source_ids":["sr-65e85a5ff56321ef9af773e89e4d758d"],"technology":"vllm","known_exclusions":[],"conditions_summary":"### Summary\n\nThe Phi-4 multimodal audio encoder (`phi4mm_audio.py` / `phi4mm_utils.py`) runs its conformer attention in **eager mode by default**: the official HF config of `microsoft/Phi-4-multimodal-instruct` does not set `use_pt_scaled_dot_product_attention`, so the default `False` applies (`phi4mm_audio.py:166`, `:876`; `phi4mm_utils.py:1669`). The eager path fully materializes the attention probability matrix of shape `[n_chunks, 16, 500, 500]` per layer — and then **stores it on the module**: `self.attn = attn` (`phi4mm_utils.py:1802`, a `torch.jit.Attribute` declared at `:1693`), keeping a live reference for every one of the 24 encoder layers.\n\nThe only gates on audio length are the server-side limits `VLLM_MAX_AUDIO_CLIP_FILESIZE_MB=25` and `VLLM_MAX_AUDIO_DECODE_DURATION_S=600` (`envs.py:83-84`) — i.e. a 10-minute clip is a **legal input by default**. A 600 s clip becomes ~60,000 mel frames, which the encoder unfolds into 500-frame chunks (`phi4mm_audio.py:996`, `max_seq_len = 500`), giving `n_chunks ≈ 120`. Each layer therefore materializes a `120×16×500×500` matrix (~1–2 GiB depending on dtype), and the module attribute keeps all completed layers' matrices alive, so GPU ","public_affected_ranges":["< 0.32.0"]}],"null_published_at":"2026-10-31T22:13:44.215Z","source_updated_at":"2026-10-28T21:15:42.000Z","source_disclosed_at":"2026-10-28T21:15:41.000Z","is_synthetic_fixture":false,"null_last_reviewed_at":null,"public_change_summary":"Dated EPSS estimate updated; this is not a new disclosure or proof of exploitation.","null_first_observed_at":"2026-10-31T21:56:08.420Z"}],"next_cursor":null,"snapshot_revision":"42594"}