vLLM is an inference and serving engine for large language models.
CVE-2026-115913
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; phi4mmutils.py:1669). The eager path fully materializes the attention probability matrix of shape [nchunks, 16, 500, 500] per layer — and then stores it on the module: self.attn = attn (phi4mmutils.py:1802, a torch.jit.Attribute declared at :1693), keeping a live reference for every one of the 24 encoder layers. The only gates on audio length are the server-side limits VLLMMAXAUDIOCLIPFILESIZEMB=25 and VLLMMAXAUDIODECODEDURATIONS=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 (phi4mmaudio.py:996, maxseqlen = 500), giving nchunks ≈ 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 memory climbs monotonically within a single request until an allocation fails.…
Part of the AI stack
This affects vLLM, in Inference & serving.
Assessment
- CVSS
- 6.5 (v3.1)
- Severity
- Medium
- EPSS
- 0.00329% chance of exploitation in the next 30 days
- Known exploited
- Not listed in CISA KEV
- Exploit availability
- Nothing found in Exploit-DB, Metasploit, Nuclei or public PoC repositories
- Published
- 26 October 2026
- Last updated
- 28 October 2026
- Affected
- pip · vllm
- Weakness
- CWE-770
- Assigner
- security-advisories@github.com
CVSS vectorCVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H
At the source
References
- https://github.com/vllm-project/vllm/commit/3d52578e13454de0bb8352cac988347cdcc89928
- https://github.com/vllm-project/vllm/pull/60614
- https://github.com/vllm-project/vllm/releases/tag/v0.32.0
- https://github.com/vllm-project/vllm/security/advisories/GHSA-8397-8qf6-f2gm
- https://nvd.nist.gov/vuln/detail/CVE-2026-115913
- https://github.com/advisories/GHSA-8397-8qf6-f2gm
Community proofs of concept
No community proof of concept yet. If you have one for a published vulnerability and the right to share it, add it below — it will appear here once reviewed.
Submit a proof of concept
Only for publicly disclosed vulnerabilities, and only material from testing you were authorised to perform. This platform does not buy, sell, or broker exploits — see earning from vulnerability research for legitimate bounty programs. Submissions are reviewed before they appear. Text only; do not include live credentials, personal data, or anything targeting systems you are not permitted to test.
Query this programmatically
Agents can read this through the MCP endpoint, or over HTTP:
curl https://cve.zerocrew.in/api/cve/CVE-2026-115913