Resource-Limit Bypass in vLLM via PyNvVideoCodec

Executive Summary

vLLM before 0.29.0 has a flaw in its PyNvVideoCodec decoder allocation where different sampler subclasses can independently increment counters. This allows unauthenticated attackers to bypass configured decoder limits by selecting various subclasses in video requests, leading to GPU memory exhaustion.

Useful 0 Not Useful 0

Detection Guidance

Detecting this vulnerability requires checking vLLM versions before 0.29.0 and monitoring GPU memory usage for unexplained exhaustion. Inspect vLLM installation with commands like 'pip show vllm' or 'vllm --version'. Check running processes for vLLM instances and compare versions against 0.29.0. Monitor GPU memory with 'nvidia-smi' or 'gpustat' for abnormal usage patterns.

Useful 0 Not Useful 0

Impact Analysis

Unauthenticated attackers can remotely trigger this to exhaust GPU memory, causing reduced safety margins, availability degradation, allocation failures, and potential system crashes in memory-constrained environments.

Useful 0 Not Useful 0

Mitigation Strategies

Upgrade vLLM to version 0.29.0 or later immediately. Apply patches from the official vLLM repository. Restrict video processing requests to trusted sources. Monitor GPU memory usage closely after updates. Consider disabling video processing features if not essential until patched.

Useful 0 Not Useful 0