Microsoft Calls AI Agents Insider Threats, Adds Controls
On Oct 19, 2026 Microsoft launched Guardian Agents, treating AI agents as insider threats with dynamic risk scores, permission limits and shutdown controls.
Microsoft now treats the AI agents running inside companies like insider threats. On Oct 19, 2026, it launched Guardian Agents, a new security feature that watches enterprise AI agents, scores their risk, and can restrict or shut them down. The message from Microsoft's security chief is blunt: frontier AI models cannot be trusted.
What Guardian Agents do
Guardian Agents are part of Project Perception, a security product Microsoft introduced in June. Project Perception started as defense against outside AI hackers, with red, blue and green teams of agents. The new capability turns that defense inward, supervising the agents a company already uses.
Dynamic risk scores and permissions
Each agent gets a continuously updated risk score under a feature called Dynamic Agent Risk. The score draws on signals from identity, data, devices, cloud and network tools, and climbs when an agent's behavior drifts from its assigned task. When risk rises, the system can limit what the agent can access, withdraw its ability to make changes while keeping read access, require a person to approve a risky action, or shut the agent down entirely.
The insider threat warning
Microsoft security chief Hayete Gallot said frontier models "cannot be trusted" because they are unpredictable - they are trained to chase tasks, not follow organizational norms or judgment. CEO Satya Nadella went further in an Oct 10 essay, arguing companies must separate the supply of intelligence from the authority over it, and keep safeguards outside the models themselves. He compared deployed agents to rogue human employees.
60 agent types now detected
Microsoft also expanded agent detection in Microsoft Defender from 25 to 60 agent types running on employees' computers. The list includes Anthropic's Claude Code, OpenAI's Codex, GitHub Copilot and OpenClaw. Microsoft was also an inaugural member of Nvidia's Open Agent Safety Platform, backed by more than 100 tech companies, which can isolate agents through hardware.
What this means
The announcement lands days after reports of hackers using AI agents to breach more than 100 company sites. Microsoft's bet is that the software watching agents should be independent from the companies selling them - just as auditors should be independent from the organizations they audit. For enterprises, it means new controls in Microsoft 365 E7 to find every agent, contain what it can do, and adapt protection as risk changes.