Sunday, 11 October 2026
Microsoft's CEO Says Treat the Model Like Someone Who Might Go Bad
Nadella’s Advice on Frontier Models Is to Stop Taking the Lab’s Word for It
On Saturday Satya Nadella published a long-form article on X arguing that organisations should treat frontier models the way a serious company treats an insider: useful, trusted with real work, and capable of error or compromise. His starting claim is a technical one. You can no longer trace a model’s behaviour to a specific code path, or tie an output to particular training data or weights — so the provider’s assurances stop being sufficient the moment an agent can reach sensitive information and take consequential action. Everything he prescribes therefore sits outside the model: identity and access rules, privilege limits, activity logging and containment boundaries kept in the surrounding system rather than in the model or the harness directing it; model diversity, so no single model is the only dependency for an outcome that matters; continuous testing and independent audits, with no model verifying its own work; tamper-proof, human-readable evidence for every meaningful action; and an authorised person able to pause or shut a model down mid-task. Chain-of-thought transparency helps, he says, but does not settle it — explanations are not consistently faithful, and putting one opaque model in charge of checking another only stacks black boxes. The reactions were the unusual part, because they came fast and from people who compete with him. Mustafa Suleyman, who runs Microsoft AI: “Superintelligence must be contained… Today, this is an engineering and governance challenge. And it isn’t new.” Box’s Aaron Levie said AI “will need to have its ‘zero trust’ era that security saw.” Chamath Palihapitiya thought “intent, verification and evidence will gate how quickly SI permeates enterprises and businesses.” Elon Musk managed “Interesting piece from CEO of Microsoft.” Nadella’s own article sits behind X’s subscriber wall, which is a small joke at the expense of a piece about readable evidence.
Source: Digg, October 10, 2026
Cisco Would Like You to Invite an Agent Into the Meeting, and It Has Brought Claude Along
Three days before that piece, Cisco spent WebexOne in Austin selling the thing Nadella is worried about. Collaborative AI agents can be invited into Webex spaces, meetings and calls to carry out multi-step work across Webex and third-party apps — Cisco’s own example is an analytics agent pulling usage figures and a presentation agent then updating the deck with them. Claude goes inside Webex, built on Claude Managed Agents, turning shared team context into analysis, recommendations and “coordinated action across approved agents”; that one is “soon”, with no date attached. Dialog, an all-new agentic harness for the Webex AI Agent platform, is built for long-horizon execution: agents that go on working on a customer’s behalf after the conversation has ended, carrying relationship context across channels and time, learning from interactions while staying inside enterprise policy. OpenAI’s dots get Webex context over MCP, and Cisco intends to put dots inside Webex as an always-on personal agent. RoomOS 27 arrives in November; most of the rest is “soon”, unstated, or — per outside coverage of the event — early 2027. “The next era of work isn’t about bolting chatbots onto yesterday’s tools,” said Jeetu Patel, Cisco’s president and chief product officer. Worth noting what Cisco put beside the agents rather than inside them: guardrails from Cisco AI Defense, and telemetry and observability from Splunk. Containment as a principle on Saturday; containment as a line item on a slide the Wednesday before.
Source: Cisco, October 7, 2026
Beijing Stopped Meta Buying It. Half a Billion Dollars Turned Up Anyway
Butterfly Effect, the Chinese company behind the AI agent Manus, said in a WeChat post on Thursday that it has raised more than $500 million — its first round since the Meta deal fell apart. Boyu Capital and IDG Capital led it; existing shareholders Tencent, HSG, formerly Sequoia China, and ZhenFund came along. The timeline is the story. Meta agreed in December 2025 to buy Manus for about $2 billion. In April 2026 Chinese authorities ordered the deal unwound, amid concern in China about AI talent and researchers going west. Manus resumed operating independently in August, and said it had to delete some user data as part of the separation. No valuation was disclosed for this round; a September report had the company in talks for $500 million at $4 billion, and at the time of the Meta deal it was reportedly past $100 million in annual recurring revenue. A government blocked the exit, the company raised the money by itself three days short of six months later, and the thing being argued over is an agent that books your travel.
Source: TechCrunch, October 8, 2026