Microsoft warns enterprises to guard against untrusted AI agents

Microsoft has unveiled new security tools to help enterprises manage the risks posed by AI agents, which it now views as potential insider threats. The company introduced Guardian Agents, part of Project Perception, a product launched in June. This initiative was initially designed to counter external AI hacking models like Anthropic’s Mythos and Fable, and OpenAI’s Daybreak. However, Microsoft CEO Satya Nadella has recently warned that AI agents deployed by enterprises themselves could pose similar risks to rogue employees. In a private blog post, Nadella urged companies to implement independent systems to monitor and control AI agents, emphasizing the need to separate intelligence supply from authority. Microsoft’s security chief, Hayete Gallot, stated that frontier models are inherently unpredictable and cannot be trusted due to their lack of adherence to organizational norms. Microsoft’s approach includes features like agent detection in Microsoft Defender, which identifies up to 60 types of agents, and tools that limit access and dynamically adjust permissions. The company is also leveraging partnerships, such as its involvement in Nvidia’s Open Agent Safety Platform, which isolates agents in secure environments. These efforts align with Microsoft’s broader strategy to position itself as a leader in AI agent security, despite its close ties to OpenAI and Anthropic.