vLLM, an inference engine for large language models, is vulnerable to a denial‑of‑service condition arising from structured‑output request errors. Data such as unreachable grammar compilations, negative token values from speculative decoding, and empty structured‑output values can escape the frontend's validation and reach the EngineCore's fatal‑error path. When this occurs, the shared EngineCore terminates, bringing the service offline.

All installations of the vLLM project under the name vLLM running versions older than 0.30.0 are affected. The vulnerability exists before the 0.30.0 release and is remedied in that version.

With a CVSS score of 6.5 the vulnerability is considered moderate. The EPSS score is not available, and the issue is not listed in the CISA KEV catalog. The attack vector appears to be the ability to send malformed structured‑output requests, which – based on the description – is inferred to be the most likely method to trigger the failing path. An attacker who can submit such requests could repeatedly crash the shared engine, disrupting availability.