Microsoft says workplace AI agents need their own security oversight

Microsoft is urging companies to treat AI agents used by employees less like trustworthy software and more like insiders with access to the business. These systems do more than answer questions: they can use company data and tools to write code, process expenses, or act inside core applications, often with broad permissions. Microsoft says frontier models pursue assigned goals without reliably understanding a company’s rules or judgment, so their behavior can shift as inputs and circumstances change—even without an outside attacker. The comparison is not that agents have bad intent; it is that access plus unpredictable behavior can turn a mistake into a security problem. Its answer is oversight outside the model: limit each agent’s access, monitor its actions continuously, and require people or automated safeguards to intervene when risk rises.