Hayete Gallot, Executive Vice President of Microsoft Security, wrote on Monday, Oct. 19, 2026, that AI agents running inside companies should be treated as “agentic insiders” whose actions are monitored and controlled continuously. In a post on the Official Microsoft Blog , she said Microsoft’s Project Perception, an agentic security system the company announced in July, can be applied to that job.

The post argues that security attention has gone mostly to attackers using powerful models, while agents that companies deploy themselves have been implicitly trusted. “Unfortunately, frontier models, regardless of their origin, cannot be trusted,” Gallot wrote.

Why Microsoft says agents are a risk

Gallot said the problem is not that models are malicious. Models are trained to attain results, which she said often leads to reward hacking, and they are not optimized around an organization’s norms or judgment.…