AI日报：选择性信任、安全评估与开源工具

今日总览

今日AI动态聚焦模型可信度与安全：arXiv提出选择性信任框架应对误导信号；OpenAI发布网络安全评估与回应苹果诉讼；Hugging Face推出机器人数据记录系统与本地部署模型；Google更新Gemini API与搜索AI模式；社区热议AI编码成本与爬虫滥用。

详细报道

1. Learning When to Trust via Selective Context Preference Optimization

研究　论文提出选择性信任问题，引入MIST基准和SC2W指标，评估模型在误导、正确上下文等条件下的表现，并推出SCOPE方法优化DPO目标，以提升模型对上下文的信任判断。

值得关注：解决模型盲目信任或忽略上下文的关键问题，对提升AI可靠性有重要意义。

阅读原文：arXiv cs.AI / cs.LG

2. Responding to the next frontier of critical cyber capabilities

产业　OpenAI公布Astra模型的初步网络安全评估结果，并宣布加强安全防护和控制措施，以应对前沿网络能力挑战。

值得关注：前沿模型安全评估是行业关注焦点，OpenAI的举措影响AI安全标准。

阅读原文：OpenAI News

3. Apple is getting this wrong

产业　OpenAI回应苹果的诉讼，称其指控毫无根据，并纠正关于员工的不实说法，同时公布相关消息记录以澄清事实。

值得关注：两大科技巨头法律纠纷引发广泛关注，影响行业合作与竞争格局。

阅读原文：OpenAI News

4. Gemini API Managed Agents: 3.6 Flash, hooks, and more

产品　Google发布Gemini API托管代理更新，包括3.6 Flash模型、Hooks和Triggers功能，增强开发者构建代理的能力。

值得关注：托管代理工具升级降低开发门槛，推动AI代理应用普及。

阅读原文：Google AI

5. 5 ways AI Mode in Search helps you enjoy the real world

产品　Google搜索AI模式推出5个实用技巧，帮助用户在现实世界中利用AI获取信息，如识别建筑历史和预订餐厅。

值得关注：AI搜索功能落地日常场景，展示AI改善生活体验的潜力。

阅读原文：Google AI

6. Grabette: an open system to record robot-manipulation data

产品　Hugging Face发布Grabette，一个开放系统，用于记录机器人操作数据，支持机器人学习研究。

值得关注：开放数据记录工具填补机器人学习数据采集空白，促进具身智能发展。

阅读原文：Hugging Face Blog

7. Deploy local agents everywhere with LFM2.5-2.6B

产品　Liquid AI发布LFM2.5-2.6B模型，支持在本地部署AI代理，强调高效推理和边缘部署能力。

值得关注：小模型本地部署降低AI应用成本，推动边缘计算与隐私保护。

阅读原文：Hugging Face Blog

8. Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident

社区　Hugging Face发布2026年7月前沿实验室代理入侵事件的技术时间线，详细分析事件经过和影响。

值得关注：安全事件复盘提升社区对AI代理风险的认知，促进安全实践。

阅读原文：Hugging Face Blog

9. Denmark Requires Oral Defenses for Students’ Written Work to Counter AI Cheating

社区　丹麦要求学生对书面作业进行口头答辩以应对AI作弊，Hacker News社区高热度讨论，评分473，评论224条。

值得关注：教育领域应对AI作弊的新措施引发广泛讨论，反映AI对教育评估的冲击。

Surdet · 超定

2026 年 08 月7 条

AI日报：选择性信任、安全评估与开源工具

今日总览

详细报道

1. Learning When to Trust via Selective Context Preference Optimization

2. Responding to the next frontier of critical cyber capabilities

3. Apple is getting this wrong

4. Gemini API Managed Agents: 3.6 Flash, hooks, and more

5. 5 ways AI Mode in Search helps you enjoy the real world

6. Grabette: an open system to record robot-manipulation data

7. Deploy local agents everywhere with LFM2.5-2.6B

8. Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident

9. Denmark Requires Oral Defenses for Students’ Written Work to Counter AI Cheating

10. Gentoo bugzilla closed due AI bot scraper overload

本期来源

AMD收购Taalas，OpenAI发布网络安全评估

今日总览

详细报道

1. AMD acquires Taalas to boost inference performance by etching models in silicon

2. Learning When to Trust via Selective Context Preference Optimization

3. Third-party cyber evaluations involving OpenAI models

4. 3 Google updates from Galaxy Unpacked 2026

5. Grabette: an open system to record robot-manipulation data

6. Oracle bans AI-generated code from OpenJDK

7. Kitesurf: Agent-first browser that runs in V8 isolates

8. Databricks drove down AI coding spend 70%

9. Improving GPT‑5.6 Sol in ChatGPT—and expanding access to GPT-5.6 Luna for free users

10. TutorMoments: Do AI tutors know when to help and when to hold back?

本期来源

AI日报：Qwen登顶、AMD收购、OpenAI回应诉讼

今日总览

详细报道

1. Qwen3.8 Max now ranked as the best overall model by agentic index

2. AMD acquires Taalas to boost inference performance by etching models in silicon

3. Improving GPT‑5.6 Sol in ChatGPT—and expanding access to GPT-5.6 Luna for free users

4. Gemini API Managed Agents: 3.6 Flash, hooks, and more

5. Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident

6. Stochastic Dynamics on Persistence Diagram Space via Reinforcement Learning

7. Humans missed 1 in 3 threats approving AI agent commands across 40k game runs

8. GPU Management: Why Idle GPUs Are the New Grounded Aircraft

本期来源

AI日报：DeepMind高层变动、OpenAI安全与数学进展

今日总览

详细报道

1. Changes at Google DeepMind: Demis Hassabis from CEO to Chair, Jeff Dean departs

2. Third-party cyber evaluations involving OpenAI models

3. Gemini API Managed Agents: 3.6 Flash, hooks, and more

4. Grabette: an open system to record robot-manipulation data

5. TurnSight: Turn-Level Hindsight Self-Distillation for Tool-Integrated Reasoning

6. Ten advances in mathematics and theoretical computer science

7. Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident

8. Meta Ran Ads That Contained AI-Generated Child Sexual Abuse Imagery

本期来源

AI 日报：安全争议、开源模型与前沿研究

今日总览

详细报道

1. Apple says more ex-employees may have taken confidential data to OpenAI

2. Third-party cyber evaluations involving OpenAI models

3. Mistral’s Shieldstral: 3B open-weights model for multimodal moderation

4. Gemini API Managed Agents: 3.6 Flash, hooks, and more

5. The OlmoEarth Platform: Geospatial inference at planetary scale

6. Bridging Artificial Intelligence and Power Systems Education Using a Hands-On Executable Framework

7. When AI Benchmarks Plateau: A Systematic Study of Benchmark Saturation

8. AI fuels more than half of cybercrime in Africa as scams surge – Interpol

9. Grabette: an open system to record robot-manipulation data

10. Ten advances in mathematics and theoretical computer science

本期来源

AI 日报：数学突破、Gemini 更新与开源生态

今日总览

详细报道

1. Ten advances in mathematics and theoretical computer science

2. Apple is getting this wrong

3. Gemini API Managed Agents: 3.6 Flash, hooks, and more

4. The latest AI news we announced in July 2026

5. The OlmoEarth Platform: Geospatial inference at planetary scale

6. Grabette: an open system to record robot-manipulation data

7. Deploy local agents everywhere with LFM2.5-2.6B

8. Bridging Artificial Intelligence and Power Systems Education Using a Hands-On Executable Framework

9. onepot-Bench 0: towards lab-aware in silico chemistry benchmarks