Microsoft says AI agents are like insider threats as ‘frontier models can’t be trusted’
techcrunch ·
Microsoft has introduced new security features, including Guardian Agents, to protect enterprises from the risks posed by AI agents. These tools are part of Project Perception, which was originally designed to simulate hacking scenarios but now addresses the concern that deployed AI agents may act as unpredictable insider threats. CEO Satya Nadella and security chief Hayete Gallot emphasize that frontier models cannot be trusted because they lack organizational judgment and may access sensitive data. Microsoft argues that companies must separate the supply of intelligence from the authority over it, rather than relying on assurances from model providers. This strategy involves monitoring agent behavior through Defender and limiting permissions based on specific tasks. The approach aligns with broader industry efforts, such as Nvidia’s Open Agent Safety Platform, to isolate and control AI actions. Although Microsoft invests in competitors like OpenAI and Anthropic, it positions its own cloud services as the necessary independent oversight layer.
techcrunch