A medium-severity vulnerability (CVE-2026-105758) in vLLM allows unauthenticated remote attackers to exhaust API-server memory on deployments serving Qwen2-VL or Qwen3-VL models by manipulating request-level max_frames and fps parameters, according to a GitHub security advisory1. Versions 0.24.0 through 0.29.x are affected; version 0.30.0 is patched. The advisory states that a prior fix capping num_frames (PR #51969) does not reach the Qwen samplers, which ignore that parameter entirely. The /tokenize and /invocations endpoints remain unauthenticated even when an API key is configured, the advisory said.
vLLM Discloses Unauthenticated DoS Flaw in Qwen Video Backends
GitHub advisory CVE-2026-105758 details a medium-severity memory-exhaustion vulnerability in vLLM affecting Qwen2-VL and Qwen3-VL deployments, patched in…
1[medium] vllm: vLLM: Qwen2-VL / Qwen3-VL video samplers bound on request-controlled max_frames, which the num_frames ceiling does not reachgithub.comHigh10·05 23:42Z
Was this useful?
The day’s vector, in your inbox.
Vector Wire Daily: one dispatch each publishing day.