How the attack works
An attacker or misbehaving agent builds a multimodal inference request containing a data URI with a video media type. Instead of a normal video file, the payload is a comma-separated list of thousands of base64-encoded JPEG frames. vLLM's video-loading code splits on commas and decodes every frame into memory without checking against the frame-count limit it enforces elsewhere. Decoding thousands of frames into numpy arrays at once can exhaust available memory and crash the server, denying service to other users.
Why it matters
A single crafted request can take down a shared inference server, disrupting every other user or agent depending on it, with no need for authentication beyond normal API access.
What you can do
- Set a hard limit on request body size and reject video data URIs with an excessive number of comma-separated segments before they reach the model server.
- Monitor outbound requests from agents for data URIs whose decoded frame count exceeds your configured sampler limit (default 32 in vLLM).
- Run vLLM inference behind resource limits (memory cgroups, container quotas) so a single request cannot exhaust host memory.
- Patch or upgrade vLLM once a fix for unbounded frame parsing is available, and verify the fix by replaying oversized payloads in a test environment.
Known benign look-alikes
- Video-understanding agents that legitimately sample more than 64 frames from a long clip and inline them as a video/jpeg data URI. Confirm the destination host is an approved inference endpoint and that the frame count matches the configured sampler.
- Benchmark, load-test or capacity-planning harnesses that deliberately push oversized multimodal payloads at a self-hosted vLLM instance.
- Security validation of the fix for this exact issue (regression tests that replay the malicious payload against a patched server).
- A client that inlines an animated sequence converted frame-by-frame to JPEG rather than uploading a container file and passing a URL.