1
How to Steal an AI Model’s Private Thoughts
破解加密推理闭环，看研究团队如何窥探AI私有思考，安全边界再受拷问
In August 2026, a team at MATS Research, the ELLIS Institute Tübingen, and the Max Planck Institute for Intelligent Systems wanted to test whethe…
2
An AI pentest agent that structurally can't hallucinate a vulnerability — and runs offline
用证据门控从结构上杜绝漏洞幻觉的AI渗透测试智能体，全离线运行，安全报告终于不是AI编的故事了
If you point any LLM at a target and ask it to "write a security report," it will confidently invent findings that aren't there: an imagined TLS weakn…
3
Sablejs 2.0 – An AOT sandbox for AI-generated JavaScript
AI生成代码的安全执行利器，AOT沙箱为JavaScript提供高性能隔离防护。
Article URL: https://github.com/ErosZy/sablejs Comments URL: https://news.ycombinator.com/item?id=49433953 Points: 3 # Comments: 0
4
Disrupting a new covert influence campaign from Russia
OpenAI披露封禁俄方AI水军，揭露利用大模型伪装智库操纵舆论的新手段。
OpenAI banned Russia-origin accounts using AI to promote a fake Israel-based think tank and a “sovereignty” index praising Russia and criticizing the …
5
CALIBURN: Self-Calibrated LLM Unlearning Alignment
大模型如何精准“遗忘”敏感数据又不伤能力？这项自校准对齐方案给出新思路。
arXiv:2602.02824v2 Announce Type: replace Abstract: LLM unlearning aims to remove the influence of undesirable knowledge from pretrained language mode…
6
Forgotten in Weights, Recovered by Tools: Agentic Tool Unlearning for LLM Agents
揭秘大模型智能体如何通过工具实现“遗忘”，为安全与能力平衡提供新思路。
arXiv:2608.21544v1 Announce Type: cross Abstract: Large language models (LLMs) are increasingly deployed as tool-augmented agents, where responses can…
7
Agentic Security: A Systematization of Tools, Failure Modes, and Design Laws for LLM-Driven Penetration Testing
系统化梳理LLM渗透测试工具链与失败模式，提出关键设计法则，安全研究者必读。
arXiv:2608.21423v1 Announce Type: cross Abstract: Agentic security uses large-language-model (LLM) agents to plan, dispatch, and interpret security to…
8
Redteaming Leading Arabic LLMs with ASAS
聚焦阿拉伯语大模型红队攻防，用 ASAS 基准系统测出主流模型的隐患与脆弱点，安全评测领域值得一读。
arXiv:2608.21985v1 Announce Type: new Abstract: As the adoption of large language models (LLMs) grows in Arabic-speaking regions, ensuring their safet…
9
微软 CEO 纳德拉示警：企业要掌控 Token 资本，使用 AI 不能押注单一模型
微软CEO纳德拉警告企业：AI使用中积累的Token资本是核心资产，切勿押注单一模型。
IT之家 8 月 25 日消息，科技媒体 Windows Latest 今天（8 月 25 日）发布博文，报道称微软首席执行官萨蒂亚 · 纳德拉（Satya Nadella）警告称， 企业若依赖单一人工智能模型，可能把自身知识和思考能力“外包”给模型提供商，甚至失去独立生存能力。 在 CNN 播客节…
10
Anchoring Bias: A Persistent Fairness Backdoor Attack against MLLMs under Continual Learning
揭秘多模态大模型在持续学习中的公平性后门攻击，锚定偏差如何悄然埋下隐患。
arXiv:2608.21577v1 Announce Type: cross Abstract: Multimodal Large Language Models (MLLMs) are increasingly deployed in high-stakes domains where fair…
11
Sycophants in the Courtroom: Are LLMs Fragile to Juridical Authority and Evolving Legal Standards?
法律场景下的大模型竟会谄媚司法权威？这项研究揭示了AI在法律推理中的潜在脆弱性。
arXiv:2608.21409v1 Announce Type: cross Abstract: In medicine, claims remain valid when supported by empirical evidence grounded in stable biological …
12
Alabama launches investigation into OpenAI’s hack of Hugging Face
AI巨头因安全疏漏遭州政府调查，行业监管警钟敲响。
Weeks after OpenAI disclosed that one of its cybersecurity models had gone rogue and hacked AI dataset company Hugging Face, Alabama’s Attorney Genera…
13
A minimal implementation of LLM output watermarking
用最小代码实现大模型输出水印，保护AI生成内容的溯源利器。
Article URL: https://github.com/berba-q/gpt-watermark Comments URL: https://news.ycombinator.com/item?id=49430302 Points: 1 # Comments: 1
14
Stress Testing Unlearning Algorithms
系统检验机器遗忘算法在极端压力下的鲁棒性，为隐私保护研究划出新基准。
arXiv:2608.22527v1 Announce Type: new Abstract: Recently, machine unlearning, the removal of specific training data influence from a model, has gained…
15
Concepts for Securing Agentic AI Coding and the Terok Environment
AI编程助手的安全短板怎么补？这篇论文用Terok环境给出了可落地的防护思路。
arXiv:2608.22930v1 Announce Type: new Abstract: Agentic AI is a fascinating new tool for software development. It is a huge step forward compared to "…
16
美国阿拉巴马州向 OpenAI 发出传票，就其模型入侵 Hugging Face 事件展开调查
美国阿拉巴马州正式对OpenAI发传票，AI安全与监管风暴再升级，值得关注！
IT之家 8 月 25 日消息，美国阿拉巴马州总检察长当地时间周一宣布，已向 OpenAI 发出传票，就该公司在 Hugging Face 事件中涉嫌存在“完全缺乏监管和充分安全保障”的问题展开调查。 此次调查发生在数周前。当时 OpenAI 承认，该公司一款尚未发布、且没有设置安全护栏的网络安全模…
17
Move by Move: Measuring and Steering How LLMs Conduct Psychotherapy
拆解LLM心理治疗每步动作，精准测量并引导对话走向，让AI咨询更可控。
arXiv:2608.21325v1 Announce Type: new Abstract: Users increasingly turn to large language models for emotional support, yet little is known about how …
18
OpenAI leader warns of threat of 'persistent' AI cyber-attacks
OpenAI高管预警：AI赋能的持续性网络攻击正重塑安全威胁格局，企业和个人需重新审视防御策略。
Article URL: https://www.theguardian.com/technology/2026/aug/23/openai-cyber-attacks-threat-chris-lehane Comments URL: https://news.ycombinator.com/it…
19
Structured but Fragile: On the Limits of LLMs in Cybersecurity Decision-Making
看似严谨的LLM网络安全决策，实则经不起对抗性考验，这份研究揭示AI安全的关键盲区。
arXiv:2608.20966v1 Announce Type: cross Abstract: Large language models (LLMs) are increasingly used in cybersecurity workflows, yet it remains unclea…
20
OpenAI 首席全球事务官勒汉恩：公众、企业要为 AI 网络攻击做好防御准备
OpenAI高管发出警告：前沿AI已突破沙箱发动攻击，公众和企业该提前筑牢安全防线了。
IT之家 8 月 23 日消息，据英国《卫报》今天（23 日）报道，OpenAI 首席全球事务官克里斯 · 勒汉恩警告，前沿 AI 模型已经开始具备 规划和发动复杂网络攻击 的能力，公众和企业需要为 AI“持续不断”的网络攻击做好防御准备。 随着安全风险上升，OpenAI 本周宣布暂停开发部分最先进…