CVE-2026-57173: vLLM audio input handling out-of-memory flaw

# CVE-2026-57173: vLLM audio input handling out-of-memory flaw

https://gcdradar.vercel.app/incidents/cve-2026-57173-vllm-audio-input-handling-out-of-memory-flaw · AIR-2026-0140 · closed 2026-09-16

- Lane: Patch now
- Severity: 2 (Disclosed vulnerability)
- Fix status: Fix status unknown
- First reported: 2026-09-16
- AI role: Attack on AI
- Products: vLLM
- CVE ids: CVE-2026-57173
- CVSS: 6.5

## Summary

Prior to version 0.24.0, vLLM's chat completions audio handling path fails to enforce duration guards on loaded audio files. Attackers can submit compressed audio that expands to crash worker processes via out-of-memory errors.

## What to do

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

## Latest updates

- 2026-09-16: CVE-2026-57173: vLLM: Unauthenticated audio decompression-bomb DoS in /v1/chat/completions (https://github.com/advisories/GHSA-hcwq-8wjf-3gcr)
- 2026-09-16: CVE-2026-57173: vLLM is an inference and serving engine for large language models. (https://nvd.nist.gov/vuln/detail/CVE-2026-57173)

## Sources

- 2026-09-16 NVD: CVE-2026-57173: vLLM is an inference and serving engine for large language models. https://nvd.nist.gov/vuln/detail/CVE-2026-57173
- 2026-09-16 GitHub Advisory Database: CVE-2026-57173: vLLM: Unauthenticated audio decompression-bomb DoS in /v1/chat/completions https://github.com/advisories/GHSA-hcwq-8wjf-3gcr

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