The vLLM project’s VideoMediaIO.load_base64() method processes video/jpeg data URLs by splitting the base64 data string on commas to extract individual JPEG frames. There is no limit enforced on the number of frames processed, allowing an attacker to craft a single API request containing thousands of frames. This causes the server to decode all frames into memory, resulting in uncontrolled resource consumption and an Out-of-Memory Denial of Service. The vulnerability is present in vLLM versions 0.8.0 and later and is reachable without authentication via the OpenAI-compatible chat completions API.

Successful exploitation results in a high-severity Denial of Service condition due to excessive memory consumption causing the server to crash or become unresponsive. There is no impact on confidentiality or integrity reported. The vulnerability can be triggered remotely without authentication, increasing the risk of disruption to services relying on the affected vLLM versions.

Patch status is not yet confirmed — check the vendor advisory for current remediation guidance. Until an official fix is available, consider implementing request size limits or input validation to restrict the number of frames processed in video/jpeg data URLs. Monitoring and rate limiting API requests may help reduce exposure to this vulnerability.