CVE-2025-62372: vLLM vulnerable to denial of service via multimodal embedding inputs

# CVE-2025-62372: vLLM vulnerable to denial of service via multimodal embedding inputs

https://gcdradar.vercel.app/incidents/cve-2025-62372-vllm-vulnerable-to-denial-of-service-via-mult · AIR-2025-0374 · closed 2025-11-20

- Lane: Patch now
- Severity: 2 (Disclosed vulnerability)
- Fix status: Fix status unknown
- First reported: 2025-11-20
- AI role: Attack on AI
- Products: vLLM
- CVE ids: CVE-2025-62372
- CVSS: 8.3

## Summary

Users could crash the vLLM inference engine serving multimodal models by supplying embedding inputs with incorrect shapes. The issue allowed unauthenticated denial of service attacks against the engine.

## What to do

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

## Latest updates

- 2025-11-21: CVE-2025-62372: vLLM is an inference and serving engine for large language models (LLMs). (https://nvd.nist.gov/vuln/detail/CVE-2025-62372)
- 2025-11-20: CVE-2025-62372: vLLM vulnerable to DoS with incorrect shape of multimodal embedding inputs (https://github.com/advisories/GHSA-pmqf-x6x8-p7qw)

## Sources

- 2025-11-20 GitHub Advisory Database: CVE-2025-62372: vLLM vulnerable to DoS with incorrect shape of multimodal embedding inputs https://github.com/advisories/GHSA-pmqf-x6x8-p7qw
- 2025-11-21 NVD: CVE-2025-62372: vLLM is an inference and serving engine for large language models (LLMs). https://nvd.nist.gov/vuln/detail/CVE-2025-62372

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