vLLM: user-supplied structured-output regex compiles with no timeout, hanging an inference worker

CVSS 8.7CVE-2026-55574AI/ML frameworks & servingcurated

Impact

The structured_outputs.regex request parameter is handed straight to the grammar compiler with no compilation timeout. The xgrammar backend applies no guard at all; the outlines backend rejects lookarounds and backreferences but does no complexity analysis, so nested quantifiers pass validation and blow up the state space. One request can pin a worker indefinitely, holding its GPU allocation while serving nothing. On a multi-tenant endpoint a handful of such requests takes the whole replica set out of service, and the stuck worker needs to be killed rather than drained gracefully.

Who can reach it

Any client permitted to send a completion request with structured outputs enabled; no authentication is described as required.

What to do

Upgrade vLLM to 0.24.0 or later and restart the serving workers. If upgrading has to wait, block or whitelist the structured_outputs.regex field at the gateway, or disable structured outputs for untrusted callers.

References

https://nvd.nist.gov/vuln/detail/CVE-2026-55574 https://github.com/vllm-project/vllm/security/advisories/GHSA-rwxx-mrjm-wc2m https://github.com/vllm-project/vllm/commit/2b3006076c5e9bc4cda9e03e3641388de3c5c286 https://github.com/vllm-project/vllm/pull/45118

Related entries

All AI/ML frameworks & serving entries

This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.