The vulnerability in vLLM allows remote clients to inject oversized parameters into multimodal processor kwargs, leading to excessive memory allocation and eventual server crashes, effectively a denial of service. The flaw permits attackers to supply image sizes, patch counts, crop counts, or frame counts that exceed normal resource limits, exhausting RAM and terminating the shared server process.

This denial of service exists in every vLLM release from version 0.7.3 up to, but not including, version 0.31.0. All deployments of the vllm project’s vllm component running those versions are affected.

The CVSS score of 8.7 reflects the high impact and remote reach of this fault, while the EPSS score of less than 1% suggests that exploitation is unlikely at the moment. Based on the description, it appears that the attack requires only API access via mm_processor_kwargs, and an exposed deployment could be targeted. The shared server could be taken down by a single malicious request.