GPU VulnDB

Database/AI/ML frameworks & serving

vLLM: attacker-supplied structured-output regex pins a CPU core and stalls the engine path

CVSS 5.3CVE-2026-73556AI/ML frameworks & servingcurated

Impact

The structured_outputs.regex parameter reaches lmformatenforcer.RegexParser without the compile timeout or the validation applied on other paths, so a single request carrying a catastrophically backtracking pattern occupies a CPU core and stalls the structured-output engine path. The GPU keeps its memory reservation and its allocation to that server while no useful work comes out of it, which on a GPU node is the expensive failure mode: the accelerator is neither serving nor free to reassign. One unauthenticated request is enough, and it can be repeated. Only deployments configured with the lm-format-enforcer structured-output backend are reachable; the advisory does not implicate the other backends.

Who can reach it

Anyone able to POST to /v1/completions on a server running the lm-format-enforcer backend. Unauthenticated where no API key is configured.

What to do

Upgrade to vLLM 0.26.0 and restart the affected servers. If an upgrade cannot happen now, switching the structured-output backend away from lm-format-enforcer removes the reachable path, or reject client-supplied regex at the gateway in front of the server.

References

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.