Separate the Supply of Intelligence From the Control Over It On October 19, Mi… · @LiaoMojo · TwKit

@LiaoMojo

Separate the Supply of Intelligence From the Control Over It On October 19, Microsoft's security EVP Hayete Gallot published a blog post titled Securing the agentic insider, and with it a new product line. Guardian Agents. A fleet of AI agents whose job is to watch the other AI agents inside your company around the clock. When risk rises, they cut permissions, demand human sign-off on consequential steps, and shut agents down if needed. The most revealing part is the list. Defender can now identify 60 kinds of local agents that employees bring into the work environment. Claude Code, Codex, OpenClaw are all named. The antivirus blocklist has a new category. Agents. Satya Nadella said the quiet part out loud earlier this month on his personal blog. He argued enterprises should stop trusting AI model vendors, even when those vendors promise their agentic products are safe. Companies need an independent system, outside the labs' control, to supervise what the labs' products do. His point, roughly put. The supply of intelligence and the control over it have to live apart. Coming from Microsoft, that lands differently. Microsoft sells agents itself. Copilot has more than 30 million corporate users. Microsoft owns about 27% of OpenAI and put $5 billion into Anthropic last November. Now it is telling customers not to trust either lab's promises. Selling the intelligence, selling the alarm system, and telling you not to trust the intelligence sellers. Three jobs, one company. It sounds like a paradox. The logic holds up, though. Gallot put it bluntly. Whether models are malicious was never the point. They simply do not recognize your rules. Agents also hold real privileges. They can touch data and call tools. When one goes off script, the damage is real. The timing is worth noting. Last week Reuters reported a Chinese-speaking hacker used agents to break into more than 100 company sites in a matter of weeks. Offense just got a cost structure that is absurdly cheap. Microsoft's answer this week is that defense has to change too. Security used to mean keeping thieves out. Now the biggest risk may be the employee you paid to let in. The industry is already moving this way. Microsoft is a founding member of Nvidia's Open Agent Safety Platform, backed by more than 100 tech companies. The idea is the same everywhere. Put agents in a monitored cage, and pull the plug when they misbehave. My read is that this closes a chapter of the AI honeymoon. For years, selling a model came with a bundled promise. Our models are safe. Trust us. This week Microsoft tore up that warranty itself. Buying intelligence and governing intelligence are now two businesses with two sellers. An auditor cannot audit itself. That sentence has finally arrived in the AI industry.

曝光 86 · 评论 2 · 点赞 3 · 书签 0