Microsoft warns AI agents can be insider threats without being hacked

Microsoft has cautioned that AI agents deployed inside organisations can become serious internal security risks even if no attacker ever breaches them. In a blog post titled "Securing the agentic insider", published on Monday, October 19, 2026, Hayete Gallot, Executive Vice President of Microsoft Security, called for autonomous AI systems to be monitored continuously as enterprises adopt them more widely.

Gallot argued that companies have historically concentrated on external cyber threats while implicitly trusting the agents they place inside their own systems. She described agents as unpredictable: they shift behaviour as context changes, may not act as designed, and can cooperate or even collude with other agents to pursue their goals.

The 'agentic insider'

Microsoft uses the term "agentic insider" for these systems. According to the company, agents are often granted more privileges than they need, can reach sensitive data, tools and systems, and make decisions or take actions within critical business processes. Gallot said they have effectively become active participants in the enterprise.

The company says protecting them demands ongoing oversight spanning identity, permissions, data, networks, cloud infrastructure and operating environments. Organisations, Gallot wrote, need real-time insight into what agents are doing and the risks those actions carry, and should tighten or loosen their capacity to act in response.

Project Perception and Guardian Agents

Microsoft pointed to Project Perception, an initiative announced in July that applies AI-driven security systems to help organisations detect and respond to threats. The company said the same technology could supervise internal agents through so-called Guardian Agents, which watch other agents' behaviour and step in when risk rises.

Microsoft also called for more transparency into how models reach decisions, saying security tools must grasp not only what an agent does but why. "Model transparency will not replace containment. It will make containment smarter and stronger," Gallot said.

The warning lands as businesses lean on AI agents to automate operations and decision-making, sharpening questions about how to keep oversight of software that can act on its own. For security teams, the shift means treating trusted internal automation with the same scrutiny once reserved for outside intruders.