vLLM audio downmixing differential enables moderation bypass

# vLLM audio downmixing differential enables moderation bypass

https://gcdradar.vercel.app/incidents/vllm-audio-downmixing-differential-enables-moderation-bypass · AIR-2026-0810 · closed 2026-04-02

- Lane: Patch now
- Severity: 2 (Disclosed vulnerability)
- Fix status: Fix status unknown
- First reported: 2026-04-02 (happened 2026-07-17)
- AI role: Attack on AI
- Products: vLLM, Librosa
- CVE ids: CVE-2026-34760
- CVSS: 5.9

## Summary

A processing differential in multi-channel audio downmixing enables hidden-input and moderation bypasses for audio models in vLLM. The vulnerability stems from a discrepancy between Librosa and standard weighting algorithms.

## What to do

No action recorded; read the sources below and the vendor's advisory.

## Latest updates

- 2026-07-17: CVE-2026-34760: vLLM: Processing differential in multi-channel audio downmixing enables hidden-input/moderation bypass for audio models (https://github.com/advisories/GHSA-6c4r-fmh3-7rh8)
- 2026-04-02: CVE-2026-34760: vLLM is an inference and serving engine for large language models (LLMs). (https://nvd.nist.gov/vuln/detail/CVE-2026-34760)

## Sources

- 2026-04-02 NVD: CVE-2026-34760: vLLM is an inference and serving engine for large language models (LLMs). https://nvd.nist.gov/vuln/detail/CVE-2026-34760
- 2026-07-17 GitHub Advisory Database: CVE-2026-34760: vLLM: Processing differential in multi-channel audio downmixing enables hidden-input/moderation bypass for audio models https://github.com/advisories/GHSA-6c4r-fmh3-7rh8

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