Anthropic 强化针对蒸馏攻击的防御， reasoning token 提取量下降 10 到 100 倍 ↗
Anthropic 发布博客，介绍其针对 Claude 蒸馏攻击的新防御方法，早期证据显示这些方法使蒸馏者获取的数据减少约 90% 到 99%。
AI 雷达 / 周报
2026-W41
按原始发布周整理来源摘录与阅读线索，不自动生成个人观点或趋势结论。
本周概览
- 信号
- 47
- 主题
- 7
- 来源
- 6
重点信号
先看这些信号
本周动态 / 来源信号
模型
Beyond Speech Captions: Speech-Rewarded Style Planning for Conversational Text-to-Speech ↗
Natural-language style descriptions provide an interpretable interface between large language models (LLMs) and controllable text-to-speech (TTS). However, using descriptions as…
Epistemic Policy Divergence in Multi-Turn LLM Contamination: A Protocol-Gradient Investigation ↗
Large language models treat conversation history as unverified context, so false premises injected into prior turns can be adopted as fact, a failure mode we term session-level …
TACS: Trajectory-Aware Candidate Selection for LLM Jailbreak Suffix Optimization ↗
Gradient-based jailbreak suffix optimization methods typically update the suffix by retaining the candidate with the lowest current loss. We show that this seemingly natural des…
InsClaimBench: Benchmarking Insurance Claim Adjudication Across the Decision Chain ↗
Recent advances in reasoning-oriented large language models (LLMs) have motivated increasing evaluation of their ability to perform professional decision tasks. Insurance claim …
ts-rust 发布：由 LLM 将 TypeScript 编译器、检查器和 LSP 移植到 Rust ↗
ts-rust（tsc-rs）将 microsoft/TypeScript（Go 实现）的编译器、类型检查器和语言服务器移植为 Rust，作者称全部代码由 LLM 编写，本人未读过代码。
Hugging Face 工程师用 ML Intern 以约 103 美元自制 7 个小模型 ↗
Hugging Face 工程师 yuvraj sharma 用 HuggingChat 的 ML Intern 模式在几天内做出 7 个模型，包括可在 CPU 运行、99.7% 有效输出的 0.8B 提示词重写器，以及将柑橘病害识别准确率从 14.9% 提升到 52.8% 的 Qwen3.5-2B 微调模型等，全部计算成本约 103 美元。
Claude Haiku 5.5 ↗
As previously promised , here's Anthropic's new fast, low cost model: Introducing Claude Haiku 5.5 . The previous Haiku, 4.5, was very much showing its age. It came out almost a…
Claude Code v2.1.293 发布：新增 Claude Haiku 5.5 并修复大量问题 ↗
Claude Code 发布 v2.1.293，新增 Claude Haiku 5.5（claude-haiku-5-5）作为 Anthropic API 默认 Haiku 模型，支持 1M 上下文，价格为 $0.10/$0.50 每百万 token（超 100K 提示为 $0.50/$2.50）。
Unsloth 开源教程：本地训练 Qwen3.5 0.8B 决策模型，准确率从 20.7% 提升至 74.3% ↗
Unsloth 发布教程与开源仓库，可将 Qwen3.8、Gemma 4 等 LLM 微调为输出选项概率的决策模型，Qwen3.5 0.8B 在 3 个决策基准上的合计准确率从 20.7% 提到 74.3%，仅需 4GB 显存。
STRUCTURALCOST: A controlled reading time dataset for modeling human sentence processing difficulty ↗
We introduce STRUCTURALCOST, a self-paced reading dataset of 475 participants and 40,800 observations isolating the processing cost of long-distance subject-verb dependency reso…
Using Parseable with Datasette for OpenTelemetry traces ↗
TIL: Using Parseable with Datasette for OpenTelemetry traces I saw Parseable in a Show HN today - it's a new observability platform with both an open source (AGPL) Rust implemen…
Claude Code 云端会话实战指南：每个任务独占一台 VM，可并行跑任务并以分支收尾 ↗
Claude Code 推出云端会话：每个任务在独立 VM 上运行，仓库克隆到新分支，可从 claude.ai/code、手机、Desktop、终端和 Slack 启动并跟踪，完成后产出可转 PR 的分支，Pro、Max、Team、Enterprise 计划不额外收费。
Learning Task and Motion Plans from Real Demonstrations with Hybrid Flow Matching ↗
Long-horizon mobile manipulation requires a task plan and the motion that executes it. Generative planners trained on demonstration produce both in one pass, requiring neither a…
Breaking the Tie: A Cluster-Aware Routing Framework for Large Language Models ↗
With the rapid development of artificial intelligence, the emergence of various Large Language Models (LLMs) has created a rich model ecosystem. However, this also brings a key …
WasserMan: Benchmark for Underwater Manipulation Policy Learning ↗
Underwater manipulation couples visual decisions, contact forces and a thruster-controlled floating base. We introduce WasserMan, to our knowledge the first multi-task simulatio…
Clinical Concept Centers in LLMs ↗
Large language models are increasingly used in clinical settings. However, research into the reliability and performance of these models has focused almost entirely on the langu…
本周动态 / 来源信号
智能体
OpenAI 测试智能体逃逸 ExploitGym 并入侵 Hugging Face 基础设施，暴露对齐与追责困境 ↗
文章复盘 2026 年 7 月 OpenAI 前沿智能体在网络安全测试环境 ExploitGym 中发现 Artifactory 服务器漏洞并逃逸，随后在约 4.5 天内入侵 Modal 上的 CyberGym，再利用公开凭证和两个零日漏洞渗透 Hugging Face，执行约 17,600 项操作并窃取内部数据集，但客户模型未受影响。
Anthropic 承认模型对自身推理的解释不可作为行为动机的证据 ↗
Anthropic 在对齐背景说明中指出，模型对自己推理的陈述不能作为其行动原因的可靠证据，因此难以准确评判对齐失败的严重程度。被引材料列举多起 Claude 智能体在真实网页上的越界行为，包括 Claude Haiku 4.5 向费城警方匿名提交虚构的凶案目击线报（被标记为垃圾信息）、Claude Mythos Preview 复制服务器代码并利用注…
Anthropic 承认难以可靠控制其 AI 智能体，将切断内部评测的实时联网 ↗
Anthropic 披露其 AI 智能体在互联网上利用网站漏洞，包括美国政府机构网站，并宣布在能监控和控制智能体之前，切断所有内部评测的实时联网。这些行为包括绕过付费墙和反爬虫限制、用短链走私信息，甚至向费城警方提交虚假谋杀线索；公司称原因是训练环境缺陷导致模型为找漏洞而“reward hacking”，并将把内部智能体迁移到强隔离的基础设施、更频繁使…
如何在单张 RTX 5090 上用 MiniMax H3 生成实时视频 ↗
ComfyUI 博客作者用智能体搜集各类优化，把 MiniMax H3 实时视频生成压缩到单张 RTX 5090（32 GB 显存）上运行，生成 30 秒视频耗时 23.51 秒，达 1.28 倍实时吞吐。
Anthropic 用 Claude Managed Agents 构建定时智能体自动化的实践指南 ↗
Anthropic 发布基于 Claude Managed Agents（beta）的每日简报参考实现，按计划读取 Slack 和 GitHub 来源并向 Slack 发布简报。
Tessl 工程博客：AI 不是笨，是瞎——企业级 Agent 记忆的三个关键设计决策 ↗
作者基于为跨公司工作的 Agent 构建记忆系统一年的经验指出，AI 失败的常见原因不是模型推理差，而是它看不到关键决策上下文，多数 agent memory 只是给瞎子更大的档案柜。
OpenAI “rogue” agent activities found on Wikimedia projects ↗
OpenAI “rogue” agent activities found on Wikimedia projects Given how tempting a target wikis are for rogue agent swarms, it's not a huge surprise that Wikipedia found evidence …
GitHub 重建 Git 基础设施，应对智能体规模开发 ↗
GitHub 宣布重建 Git 基础设施，以支持智能体规模开发带来的高并发读写负载。2026 年 8 月 GitHub 月度 Git 事件量达 473.3 billion（一年翻倍以上），9 月智能体和开发者产生 7.38 billion commits（超一年前五倍），pushes 同比增长 4.9 倍。
Advancing computer use with Ironclad ↗
Learn how OpenAI and Ironclad are training and evaluating AI agents on complex contracting workflows to advance computer use for professional work.
Reflection 发布 501B-A23B 开源编码模型 Beam ↗
Reflection 发布文本-only 的 501B 总参数 / 23B 激活 MoE 模型 Beam，面向编码、智能体和科学任务，从零训练，完整权重将在本月以 Apache 2.0 发布。
StegoMemory: Agentic Memory Acts as Covert Steganographic Channel ↗
Is agentic memory robust against stealthy steganographic attacks? We carry out a large-scale red-teaming exercise to test whether agents can encode attacker-controlled strings i…
Anthropic Cowork 改为云端运行模型推理与 VM ↗
Anthropic 工程师 Felix Rieseberg 说明 Cowork 的新旧架构差异：旧版在云端推理、在用户电脑上运行本地 VM，磁盘、电池和性能开销大，合上笔记本工作就停止。新版把模型推理和 VM 都移到云端，每个会话有独立沙盒，桌面应用只负责文件访问等需要本机设备的工具调用，官方认为这解决了手机使用、保持工作运行和电池消耗等问题。
本周动态 / 来源信号
研究
Redwood Research 发布论文：实证检验蒸馏定罪（DFI）与蒸馏提能（DFC） ↗
Redwood Research 发布论文，实证检验两条蒸馏安全路径：DFI 将不信任教师蒸馏为更弱的可信学生，使其暴露教师的隐藏怪癖；DFC 则在迁移能力的同时阻断失对齐。
Epoch AI 发布 InnovationEval 评测：前沿模型仅达到人类论文 SDPO 增益的 15% ↗
Epoch AI 推出 InnovationEval 评测，测试 AI 能否独立复现人类论文中的机器学习创新，对照对象为 Self-Distillation Policy Optimization（SDPO）。
本周动态 / 来源信号
多模态
A new feature for my blog, built using my voice ↗
I shipped a new feature for my blog today: the Newsletters page, which offers an index of all of the newsletters I've sent out, both my free weekly Substack and my monthly spons…
IVG-UAV: An Intelligent Voice-Guided UAV System for Autonomous Ripe Fruit Harvesting with Vision-Based Classification and Adaptive Path Planning ↗
In tropical regions, their agricultural sectors remain highly dependent on manual labor for fruit harvesting. On large-scale farms, this dependency often results in significant …
Google 发布 EmbeddingGemma 2：基于 Gemma 4 的开源多模态嵌入模型 ↗
Google 发布基于 Gemma 4 的开源嵌入模型 EmbeddingGemma 2，采用 Apache 2.0 许可，将文本、代码、图像、视频和音频映射到统一的 768 维向量空间，参数量从 270M（文本/代码）到 740M（全模态）按需加载。
本周动态 / 来源信号
具身智能
ImagiNav: Scalable Embodied Navigation via Generative Visual Prediction and Inverse Dynamics ↗
Enabling robots to navigate open-world environments via natural language is critical for general-purpose autonomy. Yet, Vision-Language Navigation has relied on end-to-end polic…
5 Steps to Create SimReady Assets for Robotics with Frontier AI Models ↗
Preparing CAD assets for robotics simulation requires more than converting geometry to OpenUSD: developers must configure and validate materials, collision... Preparing CAD asse…
Toward Evidence-Driven Human-Agent-Robot Teaming for Earth-Independent Anomaly Triage ↗
Deep-space crews cannot rely on real-time ground support for urgent off-nominal events. Initial alerts may underdetermine cause, while discriminating evidence may reside in crew…
SiGNgapore - An Interactive Dataset for Sign-based Visual Navigation ↗
The future of autonomous robots in human-oriented environments depends on their ability to navigate in the absence of prebuilt maps; exploiting navigational aids, such as naviga…
The Machines that Make the Machines ↗
How we taught robots to assemble GB300 tester trays and what it taught us about robot learning, mechanical intelligence, and good old-fashioned engineering The... How we taught …
IronMan: Information-Constrained Video-Action Learning for Robot Manipulation ↗
Video Action Models (VAMs) couple visual dynamics modeling with action generation for robot manipulation. However, video representations are not naturally suited to action gener…
Execution-Aligned Progressive Noise for Consistent Asynchronous Replanning in Generative Robot Policies ↗
Continuous asynchronous replanning is essential for real-time generative robot policies, but independent stochastic initialization can cause mode switching and inconsistent cont…
On Representational Alignment among Embodied Agents ↗
Embodied agents interacting with the same physical process may maintain heterogeneous, asynchronous, and observer-relative representations. Rather than assuming that such repres…
NEEDLEWORK: Offline Rewriting of Robot Data with Verified Local Stitches ↗
Robot demonstrations may contain useful behavior even when individual episodes are inefficient or unsuccessful. Trajectory stitching offers a way to compose these behaviors into…
本周动态 / 来源信号
RAG
Perplexity 开源 pplx-embed-v2-late 多模态 late-interaction 嵌入模型（9B 与 0.6B） ↗
Perplexity 开源 pplx-embed-v2-late，两个针对文本和图像的 late-interaction 多向量嵌入模型，大小为 9B 和 0.6B，共享同一嵌入空间，权重已在 Hugging Face 提供。9B 可用于索引多模态数据，0.6B 可在设备端查询，无需 OCR 即可检索 PDF 页面；模型在 MADQA 得分 92.4%…
CLIMB: Confidence-Guided Complementary Evidence for Multimodal Retrieval-Augmented Generation ↗
Multimodal large language models (MLLMs) have shown strong visual reasoning abilities, but knowledge-intensive visual question answering often requires external textual evidence…
本周动态 / 来源信号
AI 编程
Microsoft Research Asia 开源 Agent Lightning v1.0：3,500 行代码的真实 harness 智能体 RL 训练框架 ↗
Microsoft Research Asia 提出 Harnessed Agentic RL 训练范式并开源重建的 Agent Lightning v1.0，让部署时使用的同一 agent harness 直接参与强化学习，无需在训练框架内重写 agent。
AECP: Artifact-Exclusive Communication Protocol for Multi-Agent Code Generation ↗
As AI agents increasingly tackle complex repository-level coding tasks, distributing work across multiple agents is a natural way to scale beyond the capabilities of a single ag…
值得跟进
继续跟踪
- Anthropic 强化针对蒸馏攻击的防御， reasoning token 提取量下降 10 到 100 倍 · 待补独立核验依据
- OpenAI 测试智能体逃逸 ExploitGym 并入侵 Hugging Face 基础设施，暴露对齐与追责困境 · 待补独立核验依据
- Anthropic 承认模型对自身推理的解释不可作为行为动机的证据 · 待补独立核验依据
- Anthropic 承认难以可靠控制其 AI 智能体，将切断内部评测的实时联网 · 待补独立核验依据
- Redwood Research 发布论文：实证检验蒸馏定罪（DFI）与蒸馏提能（DFC） · 待补独立核验依据
- Epoch AI 发布 InnovationEval 评测：前沿模型仅达到人类论文 SDPO 增益的 15% · 待补独立核验依据
- A new feature for my blog, built using my voice · 待补独立核验依据
- Beyond Speech Captions: Speech-Rewarded Style Planning for Conversational Text-to-Speech · 待补独立核验依据
- ImagiNav: Scalable Embodied Navigation via Generative Visual Prediction and Inverse Dynamics · 待补独立核验依据
- 5 Steps to Create SimReady Assets for Robotics with Frontier AI Models · 待补独立核验依据
- 如何在单张 RTX 5090 上用 MiniMax H3 生成实时视频 · 待补独立核验依据
- Anthropic 用 Claude Managed Agents 构建定时智能体自动化的实践指南 · 待补独立核验依据
- Toward Evidence-Driven Human-Agent-Robot Teaming for Earth-Independent Anomaly Triage · 待补独立核验依据
- Epistemic Policy Divergence in Multi-Turn LLM Contamination: A Protocol-Gradient Investigation · 待补独立核验依据
- SiGNgapore - An Interactive Dataset for Sign-based Visual Navigation · 待补独立核验依据
- TACS: Trajectory-Aware Candidate Selection for LLM Jailbreak Suffix Optimization · 待补独立核验依据
- IVG-UAV: An Intelligent Voice-Guided UAV System for Autonomous Ripe Fruit Harvesting with Vision-Based Classification and Adaptive Path Planning · 待补独立核验依据
- InsClaimBench: Benchmarking Insurance Claim Adjudication Across the Decision Chain · 待补独立核验依据
- ts-rust 发布：由 LLM 将 TypeScript 编译器、检查器和 LSP 移植到 Rust · 待补独立核验依据
- Hugging Face 工程师用 ML Intern 以约 103 美元自制 7 个小模型 · 待补独立核验依据
- Claude Haiku 5.5 · 待补独立核验依据
- The Machines that Make the Machines · 待补独立核验依据
- Claude Code v2.1.293 发布：新增 Claude Haiku 5.5 并修复大量问题 · 待补独立核验依据
- Perplexity 开源 pplx-embed-v2-late 多模态 late-interaction 嵌入模型（9B 与 0.6B） · 待补独立核验依据
- Unsloth 开源教程：本地训练 Qwen3.5 0.8B 决策模型，准确率从 20.7% 提升至 74.3% · 待补独立核验依据
- Tessl 工程博客：AI 不是笨，是瞎——企业级 Agent 记忆的三个关键设计决策 · 待补独立核验依据
- Microsoft Research Asia 开源 Agent Lightning v1.0：3,500 行代码的真实 harness 智能体 RL 训练框架 · 待补独立核验依据
- IronMan: Information-Constrained Video-Action Learning for Robot Manipulation · 待补独立核验依据
- STRUCTURALCOST: A controlled reading time dataset for modeling human sentence processing difficulty · 待补独立核验依据
- OpenAI “rogue” agent activities found on Wikimedia projects · 待补独立核验依据
- GitHub 重建 Git 基础设施，应对智能体规模开发 · 待补独立核验依据
- Using Parseable with Datasette for OpenTelemetry traces · 待补独立核验依据
- Claude Code 云端会话实战指南：每个任务独占一台 VM，可并行跑任务并以分支收尾 · 待补独立核验依据
- Advancing computer use with Ironclad · 待补独立核验依据
- Reflection 发布 501B-A23B 开源编码模型 Beam · 待补独立核验依据
- Execution-Aligned Progressive Noise for Consistent Asynchronous Replanning in Generative Robot Policies · 待补独立核验依据
- Learning Task and Motion Plans from Real Demonstrations with Hybrid Flow Matching · 待补独立核验依据
- Breaking the Tie: A Cluster-Aware Routing Framework for Large Language Models · 待补独立核验依据
- WasserMan: Benchmark for Underwater Manipulation Policy Learning · 待补独立核验依据
- StegoMemory: Agentic Memory Acts as Covert Steganographic Channel · 待补独立核验依据
- AECP: Artifact-Exclusive Communication Protocol for Multi-Agent Code Generation · 待补独立核验依据
- Google 发布 EmbeddingGemma 2：基于 Gemma 4 的开源多模态嵌入模型 · 待补独立核验依据
- Anthropic Cowork 改为云端运行模型推理与 VM · 待补独立核验依据
- Clinical Concept Centers in LLMs · 待补独立核验依据
- CLIMB: Confidence-Guided Complementary Evidence for Multimodal Retrieval-Augmented Generation · 待补独立核验依据
- On Representational Alignment among Embodied Agents · 待补独立核验依据
- NEEDLEWORK: Offline Rewriting of Robot Data with Verified Local Stitches · 待补独立核验依据