CVE-2026-73559: Denial of service vulnerability in vLLM completion prompt preprocessing

# CVE-2026-73559: Denial of service vulnerability in vLLM completion prompt preprocessing

https://gcdradar.vercel.app/incidents/cve-2026-73559-denial-of-service-vulnerability-in-vllm-compl · AIR-2026-0077 · 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-2026-73559
- CVSS: 6.5

## Summary

Unbounded prompt inputs accepted by the vLLM completions endpoint between versions 0.19.0 and 0.26.0 allow authenticated API clients to exhaust server CPU and memory resources. The flaw leads to asynchronous scheduling capacity degradation.

## What to do

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

## Latest updates

- 2026-08-13: CVE-2026-73559: vLLM: Completion prompt lists fan out into unbounded engine requests (https://github.com/advisories/GHSA-87x5-vmc3-756j)
- 2026-08-13: CVE-2026-73559: vLLM is an inference and serving engine for large language models. (https://nvd.nist.gov/vuln/detail/CVE-2026-73559)

## Sources

- 2026-08-13 NVD: CVE-2026-73559: vLLM is an inference and serving engine for large language models. https://nvd.nist.gov/vuln/detail/CVE-2026-73559
- 2026-08-13 GitHub Advisory Database: CVE-2026-73559: vLLM: Completion prompt lists fan out into unbounded engine requests https://github.com/advisories/GHSA-87x5-vmc3-756j

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