CVE-2026-22773: vLLM Denial of Service via Idefics3 Vision Model Payload

# CVE-2026-22773: vLLM Denial of Service via Idefics3 Vision Model Payload

https://gcdradar.vercel.app/incidents/cve-2026-22773-vllm-denial-of-service-via-idefics3-vision-mo · AIR-2026-0432 · closed 2026-01-10

- Lane: Patch now
- Severity: 2 (Disclosed vulnerability)
- Fix status: Fix status unknown
- First reported: 2026-01-10 (happened 2026-01-13)
- AI role: Attack on AI
- Products: vLLM
- CVE ids: CVE-2026-22773
- CVSS: 6.5

## Summary

An unhandled runtime error in vLLM occurs when processing a crafted 1x1 pixel image payload for Idefics3 models. This tensor dimension mismatch causes a complete server crash.

## What to do

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

## Latest updates

- 2026-01-13: CVE-2026-22773: vLLM is vulnerable to DoS in Idefics3 vision models via image payload with ambiguous dimensions (https://github.com/advisories/GHSA-grg2-63fw-f2qr)
- 2026-01-10: CVE-2026-22773: vLLM is an inference and serving engine for large language models (LLMs). (https://nvd.nist.gov/vuln/detail/CVE-2026-22773)

## Sources

- 2026-01-10 NVD: CVE-2026-22773: vLLM is an inference and serving engine for large language models (LLMs). https://nvd.nist.gov/vuln/detail/CVE-2026-22773
- 2026-01-13 GitHub Advisory Database: CVE-2026-22773: vLLM is vulnerable to DoS in Idefics3 vision models via image payload with ambiguous dimensions https://github.com/advisories/GHSA-grg2-63fw-f2qr

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