vLLM: five more request-driven DoS and info-leak CVEs (CVE-2026-126394, 126395, 126396, 126398, 126399), fixed in 0.30.0 / 0.31.0
Alongside the Cohere2 Vision DoS (CVE-2026-126397) covered earlier, five more vLLM CVEs went public on 13 Nov 2026. Clients of the OpenAI-compatible API can exhaust memory or CPU with request-level video settings (num_frames=-1, a huge num_frames on the openpangu backend, a large MiniMax M3 min_pixels), crash the engine with an embeddings-only request on XD-RoPE models, or read live heap addresses from error messages. Upgrade to vLLM 0.31.0.
What it is
CVE-2026-126399 / GHSA-vxqj-p4gw-9h4c (6.5, 0.19.0 to before 0.31.0): media_io_kwargs.video.num_frames=-1 in a chat request overrides the server frame limit, so VideoMediaIO decodes every base64 JPEG frame and stacks them, bypassing the earlier fix for CVE-2026-34755.
CVE-2026-126394 / GHSA-894j-wg8m-63h8 (5.3, 0.30.0 to before 0.31.0): selecting video_backend "openpangu" per request on any video-capable deployment uses an unclamped num_frames, driving huge np.linspace allocations that can OOM-kill the API server without authentication.
CVE-2026-126396 / GHSA-w4cg-mhc7-mgw3 (5.3, 0.24.0 to before 0.31.0): the MiniMax M3 VL video processor accepts any min_pixels from mm_processor_kwargs, and its video-wide pixel budget only applies when max_long_side_pixel is set (default None).
CVE-2026-126398 / GHSA-5q78-j82c-2vr2 (5.9, 0.28.0 to before 0.30.0): a /v1/completions request with only prompt_embeds hits an assert in _init_xdrope_positions on XD-RoPE models (e.g. tencent/HunyuanOCR) run with --enable-prompt-embeds, terminating the engine and API server; an unpatched sibling of CVE-2026-55514.
CVE-2026-126395 / GHSA-m9fg-w9qh-67pc (5.3, 0.14.0rc1 to before 0.31.0): sanitize_message() only scrubbed lowercase hex addresses, so a crafted PNG sent to endpoints such as /tokenize returns PIL error text with live heap addresses, useful for ASLR bypass.
How to check
- Run pip show vllm (or check your image tag); anything from 0.14.0rc1 up to 0.30.x has at least one of these.
- Check whether multimodal/video input, --enable-prompt-embeds or XD-RoPE / MiniMax M3 models are enabled, and whether the API is reachable without authentication.
How to fix
- Upgrade to vLLM 0.31.0 or later.
- Until then, put the API behind authentication, disable video input you do not use, and drop media_io_kwargs and mm_processor_kwargs from client requests at your proxy.
pip show vllm | grep -i ^version
pip install -U 'vllm>=0.31.0' Sources
- criticalCVE-2026-61732: web pages can forge operator turns in the Decepticon hacking agent with ChatML tokens and run commands, upgrade to 1.1.17
- criticalCVE-2026-22778: vLLM has RCE In Video Processing
- criticalCVE-2025-47277: vLLM Allows Remote Code Execution via PyNcclPipe Communication Service
- criticalCVE-2025-32444: vLLM Vulnerable to Remote Code Execution via Mooncake Integration