CVE-2026-73557: vLLM prompt embed check bypass vulnerability

# CVE-2026-73557: vLLM prompt embed check bypass vulnerability

https://gcdradar.vercel.app/incidents/cve-2026-73557-vllm-prompt-embed-check-bypass-vulnerability · AIR-2026-0078 · closed 2026-08-13

- Lane: Patch now
- Severity: 2 (Disclosed vulnerability)
- Fix status: Fix status unknown
- First reported: 2026-08-13
- AI role: Attack on AI
- Products: vLLM
- CVE ids: CVE-2026-73557
- CVSS: 6.3

## Summary

A race condition in vLLM allows concurrent prompt embeds to bypass sparse tensor checks and reach the dense tensor conversion. The issue was addressed in version 0.26.0.

## What to do

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

## Latest updates

- 2026-09-04: CVE-2026-73557:  vLLM: Incomplete CVE-2025-62164 remediation can be bypassed by concurrent prompt parts (https://github.com/advisories/GHSA-pr7f-p5mw-fc87)
- 2026-08-13: CVE-2026-73557: vLLM is an inference and serving engine for large language models. (https://nvd.nist.gov/vuln/detail/CVE-2026-73557)

## Sources

- 2026-08-13 NVD: CVE-2026-73557: vLLM is an inference and serving engine for large language models. https://nvd.nist.gov/vuln/detail/CVE-2026-73557
- 2026-09-04 GitHub Advisory Database: CVE-2026-73557:  vLLM: Incomplete CVE-2025-62164 remediation can be bypassed by concurrent prompt parts https://github.com/advisories/GHSA-pr7f-p5mw-fc87

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