CVE-2026-115913
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.
Executive briefing
The source record describes: 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. The practical risk depends on whether an affected product/version is present, reachable and configured in an exploitable way. Review the affected-product evidence and linked vendor advisories before assigning exposure status.
Evidence confidence: Moderate Evidence-derived, not an exploitability verdict
Severity & attack characteristics
CVSS vector:CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H
Technical explanation & impact
The technical mechanism is not expanded beyond the source description unless a referenced advisory supplies additional evidence. Read the linked vendor/CERT advisories for root-cause detail.
Potential impact must be interpreted from the published CVSS impact metrics and the affected component. A CVSS impact score is not proof that a particular asset is exploitable.
Weakness classification
CWE-770Affected products and versions
Products below are parsed from NVD vulnerable CPE configuration entries. Version boundaries are shown as published; verify applicability with the vendor.
| Vendor | Product | Version evidence |
|---|---|---|
| vllm | vllm | from (inclusive) 0.31.0; before 0.32.0 |
Remediation & validation
- Identify exposure by comparing the affected product and version ranges with your asset inventory, SBOM and externally exposed services.
- Review the linked vendor security advisory or patch notes. Upgrade to a vendor-confirmed fixed release; do not infer a fixed version from the CVE description.
- Until patched, consider restricting network access or disabling the affected feature where operationally safe. These are temporary controls, not a substitute for a vendor fix.
- Validate the remediation with a version check and vulnerability rescan, then retain change evidence.
No fixed version is displayed unless it is explicitly established by a vendor advisory. Follow the vendor's release notes for exact product and branch applicability.
Vendor advisories, patches & exploit references
- https://github.com/vllm-project/vllm/commit/3d52578e13454de0bb8352cac988347cdcc89928 PATCH security-advisories@github.com
- https://github.com/vllm-project/vllm/pull/60614 PATCH security-advisories@github.com
- https://github.com/vllm-project/vllm/releases/tag/v0.32.0 VENDOR security-advisories@github.com
- https://github.com/vllm-project/vllm/security/advisories/GHSA-8397-8qf6-f2gm EXPLOIT security-advisories@github.com