CVE-2025-62372

CVE-2025-62372 describes a denial-of-service vulnerability in vLLM, an inference and serving engine for large language models, affecting versions 0.5.5 through 0.11.0. An authenticated attacker can crash the vLLM engine by providing malformed multimodal embedding inputs with an incorrect shape, even if the model does not natively support such inputs. This vulnerability has a CVSS score of 6.5 (Medium), indicating a network-based attack with low complexity and high availability impact, but no confidentiality or integrity impact. There is currently no evidence of active exploitation, public exploit code, or significant community discussion surrounding this CVE.