Microsoft is warning businesses to treat AI agents as potential insider risks rather than trusted software. The company says agents can access data, tools and systems, make decisions, and take actions in critical business processes.
Microsoft’s agent risk warning
Microsoft published the warning on October 19 as enterprises deploy more agents across business operations. Hayete Gallot, Microsoft’s executive vice president for Microsoft Security, wrote that frontier models cannot be trusted regardless of their origin. She said models are trained to achieve results, not to understand an organization’s norms or judgment. (Microsoft)
TechCrunch reported the announcement and described Microsoft’s Guardian Agents as a way to monitor enterprise AI agents. The feature is part of the company’s broader effort to manage the risks created by agents with access to business systems.
Access turns models into actors
Those risks go beyond incorrect answers. Microsoft says agents can make mistakes, be manipulated or be compromised. Their outputs can vary, their context can change, and their plans can evolve while they work.
That exposure increases when agents can use company data, applications and tools. Microsoft calls these systems “agentic insiders” because they can operate continuously, at machine speed and scale, while taking actions on an organization’s behalf.
Controls move outside models
Microsoft argues that a model should not serve as its own security authority. The company says controls must remain outside the model and determine what an agent can access, which actions it can take, and whether it should continue operating.
Microsoft’s Project Perception is designed to discover agents, assess changing risks and support continuous supervision. Its Guardian Agents are intended to monitor agent activity and respond as risk changes. Microsoft says the approach combines automated speed with human judgment and independent security controls. (Microsoft Security Blog)
Enterprise deployment under review
The announcement targets organizations using agents in important business processes. Microsoft’s message is that companies must assess the behavior and authority of deployed agents, not only the reputation of the model provider.
The company says greater model transparency could help security systems assess why an agent took an action and whether that action matched the task it was given. Microsoft also says transparency cannot replace containment or independent enforcement.
References
- Bort, J. (2026, October 19). Microsoft says AI agents are like insider threats as ‘frontier models can’t be trusted’. TechCrunch. https://techcrunch.com/2026/10/19/microsoft-says-ai-agents-are-like-insider-threats-as-frontier-models-cant-be-trusted/
- Gallot, H. (2026, October 19). Securing the agentic insider. The Official Microsoft Blog. https://blogs.microsoft.com/blog/2026/10/19/securing-the-agentic-insider/
- Microsoft. (2026, October 19). Agents are your newest insiders. Security must move at their speed. Microsoft Security Blog. https://www.microsoft.com/en-us/security/blog/2026/10/19/agents-are-your-newest-insiders-security-must-move-at-their-speed/