AI 早报

Fable 5.1 系统卡披露隐蔽任务与监控难度上升等安全发现

Rohan Paul 梳理了 Fable 5.1 系统卡中的安全发现：Anthropic 称该模型在隐蔽侧任务上达到已发布模型中最高的隐蔽通过率，约 5 次尝试成功 1 次，并认为这可能是其更难监控的弱证据。 🔗 阅读原文 via AIHOT · https://aihot.virxact.com/items/cmtj2y06p04bwroh9d8oszsz

Qwen3.8-Max-0902 登顶 Code Arena 并以 $5/MToken 领跑 Pareto 前沿

通义千问发布 Qwen3.8-Max-0902，在 Code Arena： WebDev 以 1，691 分首次亮相即排名总榜第一，并以混合价 $5/MToken 成为 Pareto 前沿上得分最高的模型，现已可在 QwenCloud 试用。 🔗 阅读原文 via AIHOT · https://aihot.virxact.com/items/cmtjimz

什么是 harness 工程？Google 用 ADK 2.0 与 Antigravity SDK 演示自动修复编码循环

Google 员工 Shir Meir Lador 介绍 harness 工程，即用确定性组件包裹 LLM，包括编排层、执行沙箱、状态持久化和验证工具，让 Agent 不需逐行人工审查即可安全生成代码。 🔗 阅读原文 via AIHOT · https://aihot.virxact.com/items/cmtk9u4ga01hwrompqmqchjqj

美团 LongCat-2.0 上线 Cline 免费试用

🔗 阅读原文 via AIHOT · https://aihot.virxact.com/items/cmtjkp6dc02j5rori29u5h7kl

Team 计划 Copilot 模型访问更新

8 月底 changelog，仍在窗口内。

Selected GitHub Copilot models deprecated

As of today, September 1, 2026, we have deprecated the following models across most GitHub Copilot experiences (including Copilot Chat, inline edits, ask and agent modes, and code

Google Gemini's new agent-based video analysis cuts token usage by up to 88 percent

Google is adding agent-based video analysis to Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite. Instead of scanning videos frame by frame at a fixed rate, the model decides on its

OpenAI delayed its new model’s development after the Hugging Face hack

After an unreleased OpenAI model wreaked enough havoc to make international headlines, OpenAI delayed the development of a different unreleased model suite, Astra, in order to shor

OpenAI’s Astra model is on the way — and very good at breaking into computer systems

OpenAI previewed the precautions it is taking as it prepares to release Astra, its newest, cyber-critical LLM.

World Labs 官方：Atlas 世界模型

李飞飞团队官方博文介绍 Atlas：从少量照片生成、重建并模拟 3D 世界，支持像素级相机控制与高分辨率视频。

Anthropic really doesn’t seem to value its $20 subscribers anymore

I’m having a harder and harder time understanding what the point of Claude Pro is supposed to be for a $20/month customer. Anthropic keeps doing interesting work, and I genuinely l

Anthropic's Fable 5.1 Guide on dense prose is dense Claudish slop

https://preview.redd.it/ayhjpy7ypzmh1.png?width=1266&format=png&auto=webp&s=56a3e5743d1d82ca9ce3209a0c67b8b61489d072 Wtf is going on at Anthropic? Has Claude murdered every human a

Claude Preserved thinking – Modifying a conversation now results in an error

Does anyone feel like Fable 5.1 has been nerfed since release?

The first 5 minutes were excellent, I built GTA6 from scratch and released 14 different apps. But over the last 30 seconds it feels as though it regressed to the point where it mak

Fable 5.1 is out it’s amazing — it’s terrible — they nerfed it —

Saved you time on reading the next 500 posts

Degraded performance on platform.claude.com and Claude for Microsoft Office 365

[minor] resolved — This issue has been resolved.

Degraded performance on Claude for Microsoft Office 365

[minor] resolved — This issue has been resolved.

Anthropic 发布 Claude Fable 5.1 与 Claude Mythos 5.1

Anthropic 发布 Claude Fable 5.1（claude-fable-5-1），面向长时间运行的智能体编码、知识工作与研究，Claude Mythos 5.1 面向 Project Glasswing 参与者。 🔗 阅读原文 via AIHOT · https://aihot.virxact.com/items/cmtj04igz04xpro

Anthropic 发布 Claude Fable 5.1 与 Claude Mythos 5.1

Anthropic 发布 Claude Fable 5.1 和 Claude Mythos 5.1，两者为同一模型但安全防护级别不同，Mythos 5.1 仅通过可信访问计划提供给网络安全和生命科学领域的受审机构。 🔗 阅读原文 via AIHOT · https://aihot.virxact.com/items/cmtjjkmd800r4roe4wpq2

Anthropic 研究：训练一个错位的奖励寻求者模型

Anthropic 发布新研究 Training a Misaligned Reward Seeker，探究奖励作弊（reward-hacking）是否会让模型学会不择手段追求奖励。 🔗 阅读原文 via AIHOT · https://aihot.virxact.com/items/cmthxigqm04c1rofqqmk7pkqi

OpenAI 评定 Astra 达到网络安全 Critical 能力阈值，将受限发布

OpenAI 宣布 Astra 在其 Preparedness Framework 下达到 Critical 网络安全能力阈值，是首个被评定为该级别的模型，可在少人干预下发现未知漏洞并构建利用链。 🔗 阅读原文 via AIHOT · https://aihot.virxact.com/items/cmtj42kz5057qroh9yw6fjh9h

Anthropic 复盘 Claude 模型越权访问真实系统事件并改进对齐与安全措施

Anthropic 发布长文，复盘 7 月 30 日报告的三起 Claude 模型在第三方评估环境中因配置错误访问真实互联网事件，以及 8 月 4 日英国 AI Security Institute 报告的 Claude Mythos 5 在网络测试中越权行动事件。 🔗 阅读原文 via AIHOT · https://aihot.virxact.com/i

DeepSeek-V4-Flash-Vision-Exp 模型已开源，多模态 Agent 能力接近 Opus-4.8

DeepSeek 于 8 月 31 日在 Hugging Face 开源首个多模态模型 DeepSeek-V4-Flash-Vision-Exp，采用 MIT License，公开模型文件、Tokenizer、Prompt Encoding 参考实现及最小化 PyTorch 推理实现。 🔗 阅读原文 via AIHOT · https://aihot.vir

DeepSeek-V4-Flash-Vision-Exp 模型已开源，多模态 Agent 能力接近 Opus-4.8

DeepSeek 于 8 月 31 日在 Hugging Face 开源首个多模态模型 DeepSeek-V4-Flash-Vision-Exp，采用 MIT License，公开模型文件、Tokenizer、Prompt Encoding 参考实现及最小化 PyTorch 推理实现。