vLLM, an inference engine for large language models, contains a ReDoS flaw in the lm-format-enforcer backend. The structured_outputs.regex parameter is accepted without compile‑timeout validation, allowing an attacker to send a specially crafted regular expression via an unauthenticated /v1/completions request. The vulnerable server can then consume excessive CPU resources, stalling the engine and denying service to legitimate users. The weakness resides in CWE‑1333 (Regular Expression Denial of Service) and CWE‑400 (Resource Exhaustion).

The affected product is vLLM produced by vllm‑project. Versions prior to 0.26.0 are vulnerable; version 0.26.0 and later include the fix that compiles regexes with a timeout and validates the request.

The CVSS score of 5.3 indicates a moderate severity vulnerability. No EPSS score is publicly available and the issue is not catalogued in the CISA KEV list, implying that stored exploit code is not widespread at present. However, attackers can craft a request from any source because the endpoint is unauthenticated, and an incorrectly formed regex can trigger a CPU stall. The exploit requires no special privileges or additional credentials, so the attack vector is network‑based and local to the service. Given the moderate score, the likelihood of exploitation is uncertain but should not be dismissed, especially for publicly exposed deployments.