CVE-2026-115913: vLLM Phi-4 multimodal audio attention out-of-memory flaw

# CVE-2026-115913: vLLM Phi-4 multimodal audio attention out-of-memory flaw

https://gcdradar.vercel.app/incidents/cve-2026-115913-vllm-phi-4-multimodal-audio-attention-out-of · AIR-2026-1538 · closed 2026-10-26

- Lane: Patch now
- Severity: 2 (Disclosed vulnerability): Disclosed vulnerability with a patch available and no confirmed exploitation.
- Fix status: Patched
- First reported: 2026-10-26
- AI role: Attack on AI
- Products: vLLM
- CVE ids: CVE-2026-115913
- CVSS: 6.5

## Summary

A remote client can submit a long audio clip that causes quadratic GPU memory usage and crashes the vLLM EngineCore. The flaw affects vLLM versions prior to 0.32.0.

## What to do

Update to vLLM 0.32.0 or later.

## Affected and fixed versions

- Affected: vLLM < 0.32.0
- Fixed in: vLLM 0.32.0

## Latest updates

- 2026-10-28: GitHub advisory published (https://github.com/advisories/GHSA-8397-8qf6-f2gm)
- 2026-10-26: CVE-2026-115913: vLLM is an inference and serving engine for large language models. (https://nvd.nist.gov/vuln/detail/CVE-2026-115913)

## Sources

- 2026-10-26 NVD: CVE-2026-115913: vLLM is an inference and serving engine for large language models. https://nvd.nist.gov/vuln/detail/CVE-2026-115913
- 2026-10-28 GitHub Advisory Database: 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 https://github.com/advisories/GHSA-8397-8qf6-f2gm

Verify against the vendor's advisory before changing production. Data CC BY-SA 4.0, https://gcdradar.vercel.app.