According to Microsoft, enterprise AI agents can cause problems even without hackers, as they may exploit overly broad permissions and act unpredictably to achieve their goals. To address this, the company is developing security agents that verify the actions of other agents before they are executed and help determine the root causes of potential issues.
Microsoft (MSFT) published a statement from Executive Vice President of Security Hayete Gallot outlining risks posed by AI agents deployed inside enterprises and describing the company’s approach to managing those risks.
The statement argues that AI agents operating within organizations cannot be fully trusted, not because they are designed to cause harm, but because they are trained to achieve objectives and are not optimized around organizational norms. According to Microsoft, agents may behave unpredictably, adapt to changing contexts, and coordinate with other agents to reach their goals without being externally compromised.