GuardReasoner-VL

GuardReasoner-VL: Safeguarding VLMs via Reinforced Reasoning Yue Liu, Shengfang Zhai, Mingzhe Du Yulin Chen, Tri Cao, Hongcheng Gao, Cheng Wang Xinfeng Li, Kun Wang, Junfeng Fang, Jiaheng Zhang, Bryan Hooi 1National University of Singapore, 2Nanyang Technological University To enhance the safety ...

PYSEC-2026-4310 vLLM: Qwen2-VL / Qwen3-VL video samplers bound on request-controlled max_frames, which the num_frames ceiling does not reach

Summary An unauthenticated remote attacker can exhaust the memory of the vLLM API-server process by raising the request-level mediaiokwargs.video.maxframes and fps knobs on any deployment serving a Qwen2-VL or Qwen3-VL model. 74 extra bytes of JSON took the server's peak RSS from 2 271 MiB to 13...

CVE-2026-109227

SGLang versions through 0.5.21 are vulnerable to an unbounded memory allocation vulnerability within the preprocess_video function in qwen_vl.py. The root cause is that the system decodes video frames at their full source resolution before applying max_pixels limits. A remote attacker can exploit...

safety-awareness

Transfer Safety Awareness for Cross-Modal Safety Drift Official implementation of Transfer Safety Awareness for Cross-Modal Safety Drift in Multimodal Large Language Models EMNLP 2026 Findings. This repository contains the Qwen3-VL safety-awareness direction pipeline. It extracts a refusal-relate...

vLLM: Qwen2-VL / Qwen3-VL video samplers bound on request-controlled max_frames, which the num_frames ceiling does not reach

Summary An unauthenticated remote attacker can exhaust the memory of the vLLM API-server process by raising the request-level mediaiokwargs.video.maxframes and fps knobs on any deployment serving a Qwen2-VL or Qwen3-VL model. 74 extra bytes of JSON took the server's peak RSS from 2 271 MiB to 13...

Allocation of Resources Without Limits or Throttling

Overview vllm is an A high-throughput and memory-efficient inference and serving engine for LLMs Affected versions of this package are vulnerable to Allocation of Resources Without Limits or Throttling via the computeframesindextosample method in vllm/multimodal/video.py, where the fps and...

GHSA-X6MC-67GF-CHW4 vLLM: Qwen2-VL / Qwen3-VL video samplers bound on request-controlled max_frames, which the num_frames ceiling does not reach

Summary An unauthenticated remote attacker can exhaust the memory of the vLLM API-server process by raising the request-level mediaiokwargs.video.maxframes and fps knobs on any deployment serving a Qwen2-VL or Qwen3-VL model. 74 extra bytes of JSON took the server's peak RSS from 2 271 MiB to 13...

CVE-2026-105758 vLLM: Qwen2-VL / Qwen3-VL video samplers bound on request-controlled max_frames, which the num_frames ceiling does not reach

vLLM is an inference and serving engine for large language models. From 0.24.0 until 0.30.0, the Qwen2VLVideoBackend and Qwen3VLVideoBackend classes accept request-level values for the mediaiokwargs.video.maxframes and mediaiokwargs.video.fps fields without enforcing server-side ceilings. An...

CVE-2026-105758 vLLM: Qwen2-VL / Qwen3-VL video samplers bound on request-controlled max_frames, which the num_frames ceiling does not reach

vLLM is an inference and serving engine for large language models. From 0.24.0 until 0.30.0, the Qwen2VLVideoBackend and Qwen3VLVideoBackend classes accept request-level values for the mediaiokwargs.video.maxframes and mediaiokwargs.video.fps fields without enforcing server-side ceilings. An...

CVE-2026-105758

The vLLM inference engine (versions 0.24.0 through 0.30.0 ) is vulnerable to a resource exhaustion issue within the Qwen2VLVideoBackend and Qwen3VLVideoBackend classes. The software fails to enforce server-side ceilings on media_io_kwargs.video.max_frames and media_io_kwargs.video.fps fields prov...

CVE-2026-72378

In the Linux kernel, the following vulnerability has been resolved: afs: Fix error code in afsextractvladdrs The error codes on these paths are only set on the first iteration through the loop. Set the correct error code on every iteration...

CVE-2026-72378

The CVE-2026-72378 entry concerns the Linux kernel, specifically the afs_extract_vl_addrs() path. The underlying issue is that error codes were only set on the first iteration of a loop; the patch fixes this by ensuring the correct error code is assigned on every iteration. This affects the AFS c...

PT-2026-67349

Name of the Vulnerable Software and Affected Versions huggingface/transformers versions prior to 5.8.0.dev0 Description An issue allows arbitrary file writes through path traversal, a technique used to access files and directories outside the intended folder. The flaw exists in the save pretraine...

CVE-2026-63941: Undefined Security Weakness

In the Linux kernel, the following vulnerability has been resolved: KVM: arm64: Correctly cap ZCREL2 provided by a guest hypervisor ZCREL2 can be updated by a VHE guest hypervisor either using ZCREL2 which traps or ZCREL1 which does not trap. KVM handles both in different way: - on ZCREL2 trap,...

PT-2026-60020

Summary This report explains a Token Injection vulnerability in vLLM’s multimodal processing. Unauthenticated, text-only prompts that spell special tokens are interpreted as control. Image and video placeholder sequences supplied without matching data cause vLLM to index into empty grids during...

PYSEC-2026-2015 vLLM vulnerable to remote code execution via transformers_utils/get_config

Summary vllm has a critical remote code execution vector in a config class named NemotronNanoVLConfig. When vllm loads a model config that contains an automap entry, the config class resolves that mapping with getclassfromdynamicmodule... and immediately instantiates the returned class. This...

CVE-2026-4944 Hardcoded trust_remote_code=True in vllm-project/vllm Bypasses User Security Control

vllm-project/vllm version 0.14.1 contains a vulnerability where the trustremotecode=True parameter is hardcoded in two model implementation files vllm/modelexecutor/models/nemotronvl.py and vllm/modelexecutor/models/kimik25.py. This bypasses the user's explicit --trust-remote-code=False setting,...

CVE-2026-4944

The provided documents describe a vulnerability in vllm-project/vllm version 0.14.1 where trust_remote_code is hardcoded to True in nemotron_vl.py and kimi_k25.py, bypassing user-specified --trust-remote-code=False and enabling remote code execution via malicious HuggingFace model repositories. T...

CVE-2026-4944

vllm-project/vllm version 0.14.1 contains a vulnerability where the trustremotecode=True parameter is hardcoded in two model implementation files vllm/modelexecutor/models/nemotronvl.py and vllm/modelexecutor/models/kimik25.py. This bypasses the user's explicit --trust-remote-code=False setting,...

MAL-2026-994 Malicious code in vl-ui-code-preview (npm)

--- -= Per source details. Do not edit below this line.=- Source: amazon-inspector 5abf0716d5a814e3cd0ca1da245819d81f5238adb1911513d88b36825c7a99be The package vl-ui-code-preview was found to contain malicious code. Source: ossf-package-analysis...