The vulnerability resides in the MiMoV2OmniMultiModalProcessor of vLLM, where attacker‑controlled image and audio strings are passed to a custom fetch routine that uses requests.get and Image.open instead of the intended MediaConnector. This bypasses the library’s allowed_media_domains and allowed_local_media_path checks, enabling the server to make outbound requests to arbitrary URLs and to read any files the vLLM process can access. The impact is a direct server‑side request forgery and potential disclosure of local files that may contain sensitive data, a weakness classified under CWE‑918.

The issue affects vllm‑project’s vLLM engine in all releases prior to version 0.26.0. No specific minor versions are listed; any build before the fix is vulnerable.

The CVSS score is 6.5, signifying moderate severity. The EPSS score is not available, and the vulnerability is not listed in the CISA KEV catalog. An attacker who can submit image or audio inputs to the processor can trigger the flaw, potentially accessing arbitrary remote resources or local files. Because the vulnerability relies on normal API usage and does not require elevated privileges, the likelihood of exploitation is considered moderate and should be treated as a significant concern for exposed services.