Microsoft says AI agents are like insider threats as ‘frontier models can’t be trusted’

What happened

Microsoft launched new security features to protect companies from risky AI agents. The most notable one is Guardian Agents, which watches the AI agents a company uses. Microsoft's leaders now warn that an AI agent could act like a rogue employee, which is called an insider threat.

Key facts

• Guardian Agents is part of Project Perception, which was introduced in June.
• Microsoft's Defender product can now find up to 60 types of agents that employees bring into work.
• These include Claude Code, Codex, and OpenClaw.
• Microsoft was an inaugural member of Nvidia's new Open Agent Safety Platform.

Why the change in thinking

At first, Project Perception was built to guard against outside hackers. It had a red team that attacks, a blue team that finds weak spots, and a green team that fixes them. Now Microsoft worries the danger may come from the agents companies already pay for.

Why it matters

Microsoft says AI models can be unpredictable. They are trained to finish tasks, not to follow company rules or good judgment. Nadella argues that companies should not rely only on AI labs for safety. Instead, a separate system should watch and control the agents, like auditors check the companies they audit. Microsoft also gains business by selling these tools, even though it is close to OpenAI and Anthropic.

What to watch

• Whether other companies adopt separate systems to watch AI agents.
• How Nvidia's Open Agent Safety Platform grows with industry support.
• Whether features that ask humans to approve risky agent actions become common.