CVE-2026-73556: vLLM denial of service via regex structured output

# CVE-2026-73556: vLLM denial of service via regex structured output

https://gcdradar.vercel.app/incidents/cve-2026-73556-vllm-denial-of-service-via-regex-structured-o · AIR-2026-0079 · 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-2026-73556
- CVSS: 5.3

## Summary

An unauthenticated request in vLLM can supply a catastrophic regular expression that stalls the structured-output engine and consumes a CPU core. A patch was released 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-2026-73556: vLLM: ReDoS via structured_outputs.regex in the lm-format-enforcer backend (no compile timeout) — missed sibling of GHSA-rwxx-mrjm-wc2m (https://github.com/advisories/GHSA-48jh-3gj7-fg8v)
- 2026-08-13: CVE-2026-73556: vLLM is an inference and serving engine for large language models. (https://nvd.nist.gov/vuln/detail/CVE-2026-73556)

## Sources

- 2026-08-13 NVD: CVE-2026-73556: vLLM is an inference and serving engine for large language models. https://nvd.nist.gov/vuln/detail/CVE-2026-73556
- 2026-09-04 GitHub Advisory Database: CVE-2026-73556: vLLM: ReDoS via structured_outputs.regex in the lm-format-enforcer backend (no compile timeout) — missed sibling of GHSA-rwxx-mrjm-wc2m https://github.com/advisories/GHSA-48jh-3gj7-fg8v

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