CVE-2025-29770: vLLM denial of service via outlines unbounded cache

# CVE-2025-29770: vLLM denial of service via outlines unbounded cache

https://gcdradar.vercel.app/incidents/cve-2025-29770-vllm-denial-of-service-via-outlines-unbounded · AIR-2025-0207 · closed 2025-03-19

- Lane: Patch now
- Severity: 2 (Disclosed vulnerability)
- Fix status: Fix status unknown
- First reported: 2025-03-19
- AI role: Attack on AI
- Products: vLLM, outlines
- CVE ids: CVE-2025-29770
- CVSS: 6.5

## Summary

An unbounded grammar cache used for structured outputs in vLLM can lead to denial of service conditions. A vulnerability advisory was published detailing the risk.

## What to do

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

## Latest updates

- 2025-03-19: CVE-2025-29770: vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs. (https://nvd.nist.gov/vuln/detail/CVE-2025-29770)
- 2025-03-19: CVE-2025-29770: vLLM denial of service via outlines unbounded cache on disk (https://github.com/advisories/GHSA-mgrm-fgjv-mhv8)

## Sources

- 2025-03-19 GitHub Advisory Database: CVE-2025-29770: vLLM denial of service via outlines unbounded cache on disk https://github.com/advisories/GHSA-mgrm-fgjv-mhv8
- 2025-03-19 NVD: CVE-2025-29770: vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs. https://nvd.nist.gov/vuln/detail/CVE-2025-29770

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