The vllm-project's vllm inference and serving engine for large language models contains a SSRF vulnerability identified as CWE-918. Specifically, versions before 0.26.0 have a flaw in the MiMoV2OmniMultiModalProcessor component where image and audio strings controlled by an attacker are passed through _fetch_image, requests.get, and Image.open instead of the intended MediaConnector. This bypasses protections such as allowed_media_domains and allowed_local_media_path, enabling server-side requests and arbitrary file reads accessible to the vLLM process. The issue is resolved in version 0.26.0.

An attacker with the ability to supply image or audio input strings can exploit this vulnerability to make server-side requests or read arbitrary files accessible to the vLLM process. This can lead to unauthorized disclosure of sensitive data. The CVSS score of 6.5 reflects a medium severity with high confidentiality impact but no integrity or availability impact.

Upgrade to vllm version 0.26.0 or later, where this SSRF vulnerability has been fixed. No additional mitigation steps are indicated as the fix addresses the root cause.