vLLM is a high-throughput inference engine for large language models. Prior to version 0.24.0, the structured_outputs.regex API accepted user-supplied regular expressions that were forwarded directly to the grammar compiler backends without any compilation timeout or complexity analysis. Malicious patterns that use nested quantifiers can trigger exponential state-space expansion, causing the inference worker to hang indefinitely and effectively deny service to other requests. The vulnerability is a classic regular expression denial of service (CWE-1333).

The flaw affects all releases of the vllm engine released by the vllm-project before version 0.24.0. Versions 0.24.0 and later address the issue by adding a timeout guard and performing complexity validation on the regex, thereby preventing unbounded compilation and restoring stability.

The CVSS score of 8.7 denotes a high severity attack. The EPSS score is below 1% and the flaw is not listed in the CISA KEV catalog, indicating a low likelihood of public exploitation at present. An attacker can trigger the denial of service by submitting a crafted regex to the structured_outputs.regex endpoint. The description does not specify authentication requirements or the exact exposure surface, so it is inferred that the vulnerability can be accessed remotely through any interface that exposes the API without adequate input validation.