safety-awareness

跨模态安全漂移的迁移安全意识 跨模态安全漂移中多模态大语言模型的迁移安全意识 （EMNLP 2026 Findings）的官方实现。 本仓库包含 Qwen3-VL 安全意识方向流水线。它从有害和无害多模态示例中提取与拒答相关的激活方向，选择能够改善拒答行为同时限制良性任务漂移的方向，并可选地训练一个方向特定的安全意识向量。 当前代码通过 Hugging Face Transformers 支持 Qwen/Qwen3-VL-8B-Instruct。 仓库结构 . ├── artifacts/directions/ 预计算的选定方向及元数据 ├── data/csv/ 流水线使用的 CSV ...

Sleeping Secrets: How Fine-Tuning Reawakens Privacy Risks in Language Models

Beyond adapting Large Language Models LLMs to specialized applications, fine-tuning has recently been shown to recover private information that is no longer accessible through direct queries. Previous fine-tuning recovery attacks, however, require genuine private supervision drawn from the same...

vLLM has Remote DoS via Invalid Recovered Token Reinjection

SummaryA frontend-legal multi-request speculative workload can make vLLM produce an out-of-vocabulary recovered token equal to vocabsize, convert that value to -1 when choosing the next live token for a request, and then feed that -1 back into the next drafter input ids. On Qwen3 GPTQ this reache...

PYSEC-2026-2014 vLLM has remote code execution vulnerability in the tool call parser for Qwen3-Coder

Summary An unsafe deserialization vulnerability allows any authenticated user to execute arbitrary code on the server if they are able to get the model to pass the code as an argument to a tool call. Details vLLM's Qwen3 Coder tool parser contains a code execution path that uses Python's eval...

Maestro 0.15.4

Maestro is a cross-platform desktop app for orchestrating your fleet of AI agents and projects. It's a high-velocity solution for hackers who are juggling multiple projects in parallel. Designed for power users who live on the keyboard and rarely touch the mouse. Collaborate with AI to create...

EUVD-2025-29456

Malicious code in bioql PyPI...

GHSA-79J6-G2M3-JGFW vLLM has remote code execution vulnerability in the tool call parser for Qwen3-Coder

Summary An unsafe deserialization vulnerability allows any authenticated user to execute arbitrary code on the server if they are able to get the model to pass the code as an argument to a tool call. Details vLLM's Qwen3 Coder tool parser contains a code execution path that uses Python's eval...

CVE-2025-9141

A vulnerability was found in vLLM's Qwen3 Coder tool parser. Since this parser uses Python's eval function, it poses a risk of arbitrary code execution. This vulnerability appears during the parameter conversion process when the parser attempts to handle complex data types...