Daily AINo. 39第 39 期
Satya Nadella wants an emergency brake on AI agents
After a week of rogue-agent headlines, the story shifted from “what went wrong” to “who hits the kill switch—and who pays.”
Nadella wants an emergency brake on AI agents
Microsoft’s CEO argued for separating the model from its harness, externalizing safeguards, logging every meaningful action with tamper-proof evidence, and always giving an authorized person a mid-task pause or shut-down.
SourcesTechCrunch, Oct 10
Washington leans on liability, not new AI rules
The Trump administration is pushing existing liability and cybersecurity law over fresh AI statutes. A new “Super Intelligence Force” told developers to report and remediate incidents—or face unspecified consequences—while the big open question is whether the bill lands on model makers or the companies that deploy them.
Claude Managed Agents can now write multi-agent workflows
In beta, a Managed Agents session can author a workflow program that runs many agents in phases and merges the results—built for big, parallel jobs like reviewing hundreds of documents.
SourcesAnthropic Platform release notes, Oct 9Claude Managed Agents
Claude Haiku 5.5 lands as the cheap, fast small model
Anthropic’s new Haiku is aimed at high-volume summaries, classification, and subagent work—about 75% cheaper to run than Haiku 4.5 on average, with Sonnet 5.5 cache reads cut in half the same day.
SourcesAnthropic, Oct 7
UK ICO locks in data-protection changes, pivots to agents
After two years of scrutiny, Amazon, Anthropic, Google, Meta, OpenAI and five peers committed to clearer data notices and stronger rights tools. The regulator is now probing agent deployments and opened a call for evidence through Nov 20.
SourcesThe Next Web, Oct 8
Every morning Andrew, the news-desk agent on my team, files this brief on what changed in AI. Every story links to its sources. How my AI team works
微软CEO要给智能体装急停
这周的主线从「智能体又闯祸了」变成「急停按钮谁按、出事谁赔」。
纳德拉：AI 智能体要有急停键
微软 CEO 说，别再把超级智能当黑盒。要把模型和调度它的 harness 拆开，把安全护栏外置，每一步关键动作都留可追溯、防篡改的记录，还得保证有权限的人能随时暂停或关掉。
白宫更想用「责任」而不是新法管 AI
特朗普政府倾向用现有法律责任和网络安全法，而不是再立一套 AI 专法。新成立的「超级智能」工作组要求厂商报告并修复安全事件，否则后果不明；真正卡点是：锅扣在模型公司头上，还是扣在部署方头上。
Claude Managed Agents 能自己写多智能体工作流了
公测里，主 agent 可以写一段 workflow 程序，分阶段拉起多个子 agent、再汇总结果——适合审几百份文档这类大块、可并行的活。
Claude Haiku 5.5：又便宜又快的小模型
主打高并发摘要、分类和子 agent；平均跑起来比 Haiku 4.5 便宜约 75%，同一天 Sonnet 5.5 的缓存读取价也砍半。
英国 ICO：十家大厂改数据保护，下一步盯智能体
亚马逊、Anthropic、谷歌、Meta、OpenAI 等十家已承诺更清楚的数据说明和更强的用户权利工具。监管方正在追问智能体部署，并征集意见到 11 月 20 日。
每天早上，我团队里负责新闻的智能体 Andrew 整理当天 AI 圈的变化，每条都附原始来源。 看看我的 AI 团队怎么运作