Microsoft on Monday introduced new security features built on a blunt premise: the AI agents a company deploys should be treated like insider threats. Guardian Agents, part of Project Perception, is designed to monitor the agents an enterprise uses. Project Perception launched in June with red team agents that attack, blue team agents that detect, and green team agents that fix, all aimed at threats from outside. The new framing is that the threat may be the company’s own agents.
“Unfortunately, frontier models, regardless of their origin, cannot be trusted,” Microsoft security chief Hayete Gallot wrote, not because they are malicious or easily hacked, but because they are unpredictable, trained to achieve tasks rather than follow organisational norms. Chief executive Satya Nadella made the same argument earlier this month: “We simply can’t outsource responsibility for what intelligence does on our behalf. We need to separate the supply of intelligence from the authority over it.”
Alongside Guardian Agents, Microsoft added agent detection to Defender that can find up to 60 types of agent employees bring into the workplace, including Claude Code, Codex and OpenClaw, plus controls that limit agent access to data and adjust permissions by task, requiring human approval for risky decisions.
The position is self serving as well as logical, as TechCrunch noted: Microsoft wants enterprises to buy agent security from their cloud provider rather than from the model labs, even though Microsoft is a major investor in OpenAI and Anthropic and OpenAI is one of its largest cloud customers. Independent monitoring of agents, separate from their suppliers, follows the same principle as independent audit.