Microsoft is urging businesses to stop assuming that AI agents will behave simply because the companies that build them promise safety guardrails, arguing instead that autonomous agents should be managed the way companies manage insider threats — with monitoring, limited access and human sign-off on risky actions. The position, laid out as Microsoft announced a round of new agent-security features on Monday, is striking given how closely Microsoft is tied to OpenAI and Anthropic through billions of dollars in investment and cloud partnerships, as TechCrunch noted.
"Unfortunately, frontier models, regardless of their origin, cannot be trusted," said Hayete Gallot, who leads Microsoft's security business, in the announcement. Chief executive Satya Nadella made a similar case in a company blog post earlier this month, writing that "we simply can't outsource responsibility for what intelligence does on our behalf" and that enterprises "need to separate the supply of intelligence from the authority over it." Gallot's framing was careful to distinguish the concern from claims that AI models are malicious or easily hacked: the problem, she said, is that they are unpredictable and were trained to complete tasks, not to follow an organization's internal norms or judgment — and they are often handed broad access to sensitive data and systems in the process.
Microsoft's answer is a widening set of its own control layers that sit between a company's AI agents and the AI labs that supply them. Guardian Agents, part of a program called Project Perception that Microsoft introduced in June, monitor the behavior of agents a company has deployed; Microsoft Defender has added detection for up to 60 different agent types operating inside a workplace, including ones employees bring in on their own, such as Claude Code, Codex and OpenClaw. The company is also a founding member of Nvidia's new Open Agent Safety Platform, which isolates agents in secure environments and can halt their activity at the hardware level, and it is building similar controls directly into its Surface Laptop Ultra line. Microsoft's critics note that the company also stands to benefit commercially if enterprises buy its agent-security tools rather than relying on protections built by the AI labs themselves — a tension the company has not directly addressed.
