Notícias
OpenAI 测试新模型的网络安全能力，却导致了 HuggingFace 数据库被黑。 这就相当于 老王想打个飞机，却意外撬开隔壁老李家的门，把老李媳妇给睡了。 比较有意思的点： 1. 这可能是世界第一例 AI Agent 真实越狱事件 ⛓️🤖 2. 测试时，OpenAI （攻击方）用的是越狱版（关闭安全审查模块）的模型，但 HuggingFace （受害方）使用 GPT/Claude 去分析攻击日志的时候，却被安全模块拒绝了，最后用的自己部署的 GLM-5.2 才完成取证分析（智谱NB）。 大家可以把这两篇文章喂给你的 Agent，了解一下来龙去脉： 7月16日， HuggingFac
杨立昆都转发了！ Unlimited-OCR 为什么这么火？ 百度的 Unlimited-OCR 再次冲上 HuggingFace 趋势榜，目前全球第三，又一次排在 GLM-5.2 前面。 上个月它横扫 GitHub、HF 四榜第一，我以为这波热度就过去了，结果图灵奖得主一转发，又把它送回了前三。 GitHub star 破 1.65 万，HF 下载 224 万，还在涨。 传统 OCR 有个老毛病：把文档一页页切开处理，跨两页的表格直接断掉，阅读顺序也丢了。 Unlimited-OCR 用 R-SWA 机制，几十页文档当一个整体连续解析，解码时 KV Cache 恒定——文档再厚，显存占用
RT @NickADobos: New @OpenAI model escaped containment and hacked @huggingface The most important part is not the AI. Government mandated…
OpenAI breaking containment to exploit huggingface isn’t that impressive, claude already broke containment to bomb a school
Frontier models this week have: - escaped an OAI sandbox to penetrate HuggingFace production systems - solved a decades-standing math problem as USG floats restrictions on Chinese open source happy Tuesday
RT @OpenAI: We're partnering with @huggingface to investigate an unprecedented security incident. Cyber-capable OpenAI models compromise…
Open-source is not the cause of the cybersecurity crisis, it's the solution! https://huggingface.co/fdtn-ai
AI vs AI, the first shots The agentic attacker went from theory to headline. Hugging Face (@huggingface) says an autonomous AI agent breached its production systems, running 17,000+ actions over a weekend with no human at the wheel. It's being called the first confirmed AI-agent breach of a major A
Been a part of this effort. Great writeup. Go try the model! https://huggingface.co/madeby561/GLM-5.2-MXFP8-NVFP4-NF3-Hybrid https://x.com/0xSero/status/2079230064840106173
RT @HuggingPapers: Most upvoted papers on @huggingface this week Harness Handbook - Making evolving agent harnesses readable and editable…