Microsoft launches Guardian Agents to monitor AI agents as potential internal threats

Under Project Perception, Microsoft added Guardian Agents to monitor AI agents deployed in enterprises. Launched in June, Project Perception initially included red, blue and green team agents for attacking, detecting and fixing vulnerabilities, premised on AI attackers coming from outside the enterprise. CEO Nadella recently reversed course, saying agents that enterprises pay to deploy may themselves be threats, akin to insider risk, and arguing that intelligence supply should be separated from authorization of intelligence. Security chief Gallot said frontier models cannot be trusted regardless of origin because they are unpredictable and often granted access to sensitive data and tools. Besides Guardian Agents, Microsoft added agent detection to Microsoft Defender that can find up to 60 agents employees bring to work, including Claude Code, Codex and OpenClaw. It also added the ability to restrict agent access to data and systems, dynamically adjust agent permissions by task, and require human approval for high-risk agent decisions. For enterprise security teams, this means monitoring now extends from external attackers to third-party agents in internal use, and the monitoring system must be independent of model vendors such as OpenAI and Anthropic, giving enterprises a cloud-vendor option when buying agent security products.