Microsoft sets out security controls for AI agents inside companies
Microsoft describes Guardian Agents that supervise other agents through Project Perception. Its security chief argues that frontier models are unpredictable even without being compromised; TechCrunch reports controls that restrict data access and require human approval for risky decisions.
ThereforeSecurity teams can set limits on agents separately from the companies supplying their models.
Microsoft's security announcement · TechCrunch's agent-security coverage