Resource Exhaustion in vLLM via Malicious Video Frame Requests

Executive Summary

vLLM versions from 0.19.0 to 0.31.0 have a resource exhaustion flaw. Attackers can send chat completion requests with media_io_kwargs.video.num_frames set to -1 or a large value. This forces VideoMediaIO to decode and stack many base64 JPEG frames, consuming excessive memory.

Useful 0 Not Useful 0

Detection Guidance

Monitor for unusually high memory usage during video processing in vLLM instances. Check server logs for chat completion requests with media_io_kwargs.video.num_frames set to -1 or abnormally large values. Inspect network traffic for base64-encoded JPEG frames in inference requests.

Useful 0 Not Useful 0

Impact Analysis

An attacker could exploit this to crash the vLLM server by exhausting its memory, leading to denial-of-service. Systems relying on vLLM for inference may become unavailable, disrupting services that depend on it.

Useful 0 Not Useful 0

Mitigation Strategies

Upgrade vLLM to version 0.31.0 or later. Implement input validation to reject requests with num_frames=-1 or excessively large values. Configure rate limiting for chat completion endpoints to prevent resource exhaustion attacks.

Useful 0 Not Useful 0