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 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.
Fixes and workarounds
Fix
vllm (PyPI): OSV reports a fixed version 0.32.0. Check the advisory for supported release branches.
Source: OSV GHSA-8397-8qf6-f2gmAffected products and versions
| Product | Versions | Authority |
|---|---|---|
| vllm vllm-project |
< 0.32.0 (affected) | CNA: GitHub_M |
vllmvllm
PyPI |
0.0.1 (affected, PyPI) 0.1.0 (affected, PyPI) 0.1.1 (affected, PyPI) 0.1.2 (affected, PyPI) 0.1.3 (affected, PyPI) 0.1.4 (affected, PyPI) 0.1.5 (affected, PyPI) 0.1.6 (affected, PyPI) 0.1.7 (affected, PyPI) 0.10.0 (affected, PyPI) 0.10.1 (affected, PyPI) 0.10.1.1 (affected, PyPI) 0.10.2 (affected, PyPI) 0.11.0 (affected, PyPI) 0.11.1 (affected, PyPI) 0.11.2 (affected, PyPI) 0.12.0 (affected, PyPI) 0.13.0 (affected, PyPI) 0.14.0 (affected, PyPI) 0.14.1 (affected, PyPI) 0.15.0 (affected, PyPI) 0.15.1 (affected, PyPI) 0.16.0 (affected, PyPI) 0.17.0 (affected, PyPI) 0.17.1 (affected, PyPI) 0.18.0 (affected, PyPI) 0.18.1 (affected, PyPI) 0.19.0 (affected, PyPI) 0.19.1 (affected, PyPI) 0.2.0 (affected, PyPI) 0.2.1 (affected, PyPI) 0.2.1.post1 (affected, PyPI) 0.2.2 (affected, PyPI) 0.2.3 (affected, PyPI) 0.2.4 (affected, PyPI) 0.2.5 (affected, PyPI) 0.2.6 (affected, PyPI) 0.2.7 (affected, PyPI) 0.20.0 (affected, PyPI) 0.20.1 (affected, PyPI) 0.20.2 (affected, PyPI) 0.21.0 (affected, PyPI) 0.22.0 (affected, PyPI) 0.22.1 (affected, PyPI) 0.23.0 (affected, PyPI) 0.24.0 (affected, PyPI) 0.25.0 (affected, PyPI) 0.25.1 (affected, PyPI) 0.26.0 (affected, PyPI) 0.27.0 (affected, PyPI) 0.27.1 (affected, PyPI) 0.28.0 (affected, PyPI) 0.29.0 (affected, PyPI) 0.3.0 (affected, PyPI) 0.3.1 (affected, PyPI) 0.3.2 (affected, PyPI) 0.3.3 (affected, PyPI) 0.30.0 (affected, PyPI) 0.31.0 (affected, PyPI) 0.4.0 (affected, PyPI) 0.4.0.post1 (affected, PyPI) 0.4.1 (affected, PyPI) 0.4.2 (affected, PyPI) 0.4.3 (affected, PyPI) 0.5.0 (affected, PyPI) 0.5.0.post1 (affected, PyPI) 0.5.1 (affected, PyPI) 0.5.2 (affected, PyPI) 0.5.3 (affected, PyPI) 0.5.3.post1 (affected, PyPI) 0.5.4 (affected, PyPI) 0.5.5 (affected, PyPI) 0.6.0 (affected, PyPI) 0.6.1 (affected, PyPI) 0.6.1.post1 (affected, PyPI) 0.6.1.post2 (affected, PyPI) 0.6.2 (affected, PyPI) 0.6.3 (affected, PyPI) 0.6.3.post1 (affected, PyPI) 0.6.4 (affected, PyPI) 0.6.4.post1 (affected, PyPI) 0.6.5 (affected, PyPI) 0.6.6 (affected, PyPI) 0.6.6.post1 (affected, PyPI) 0.7.0 (affected, PyPI) 0.7.1 (affected, PyPI) 0.7.2 (affected, PyPI) 0.7.3 (affected, PyPI) 0.8.0 (affected, PyPI) 0.8.1 (affected, PyPI) 0.8.2 (affected, PyPI) 0.8.3 (affected, PyPI) 0.8.4 (affected, PyPI) 0.8.5 (affected, PyPI) 0.8.5.post1 (affected, PyPI) 0.9.0 (affected, PyPI) 0.9.0.1 (affected, PyPI) 0.9.1 (affected, PyPI) 0.9.2 (affected, PyPI) Ecosystem version events: introduced 0; fixed 0.32.0 Source events (ECOSYSTEM)
Other versions: unaffected |
OSV GHSA-8397-8qf6-f2gm |
Version ranges are source assertions. Check the vendor advisory and your installed configuration before determining exposure.
CVSS assessments
| Version / score | Vector | Authority |
|---|---|---|
| 3.1 / 6.5 medium |
CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H | CNA: GitHub_M |
Exploit evidence
No exploit evidence in collected sources. This is not proof of no exploitation.
References
- https://github.com/vllm-project/vllm/security/advisories/GHSA-8397-8qf6-f2gmCNA: GitHub_M · x_refsource_CONFIRM
- https://github.com/vllm-project/vllm/pull/60614CNA: GitHub_M · x_refsource_MISC
- https://github.com/vllm-project/vllm/commit/3d52578e13454de0bb8352cac988347cdcc89928CNA: GitHub_M · x_refsource_MISC
- https://github.com/vllm-project/vllm/releases/tag/v0.32.0CNA: GitHub_M · x_refsource_MISC
- https://github.com/advisories/GHSA-8397-8qf6-f2gmOSV / GitHub Security Advisories · advisory
- https://osv.dev/vulnerability/GHSA-8397-8qf6-f2gmOSV / GitHub Security Advisories · osv
For agents and automation
Read public disclosures for free as Markdown or JSON. Compatible browsers also expose native WebMCP tools to search disclosures, read this page, and discover API access.
Use the ontology-linked Cyber API with x402 pay per request, or a Y2 Pro or Elite API key with the intel:cyber scope.
Endpoints and access details
https://api.y2.dev/api/v2/cyber/cveshttps://api.y2.dev/api/v2/cyber/cves/{cveId}
A workspace owner or admin can create a Pro or Elite API key. Send Authorization: Bearer <Y2_API_KEY> from your trusted agent runtime.
For x402, omit Authorization and read the HTTP 402 payment challenge before signing a retry. Public browser tools do not make payments.