Microsoft announces supervision over AI agents. "You cannot trust."

The corporation wants to limit this risk, among other things, using Guardian Agents, which are to constantly supervise the actions of other AI agents. The solution is part of Project Perception, a security system being developed by Microsoft. AI Agents are unpredictable "Advanced models, regardless of their origin, cannot be trusted" - wrote Hayete Gallot, head of Microsoft Security. As she explained, this does not mean that the models are inherently malicious. The problem is that they were trained to achieve results, not to follow the rules or judgment applicable in a specific organization. "Agents are unpredictable" - she emphasized. Microsoft points out that corporate agents often have broad permissions, access to data and tools, and the ability to make decisions in important processes. Project Perception is to allow for their detection and control, including assigning them identities and determining their permissions. Guardian Agents, in turn, will continuously supervise their actions and react when the risk level increases. The Need for Limited Trust Satya Nadella also wrote about the need for limited trust in AI models. "We cannot outsource responsibility for what intelligence does on our behalf. The supplier's assurances do not release us from this responsibility" - stated the Microsoft CEO. In his opinion, companies should separate the provision of models from the control over what the AI has access to and what actions it can perform. Microsoft announced Project Perception in July. The system uses teams of agents called red, blue, and green, which support the testing of security, threat detection, and response, respectively. Now Microsoft wants to use the same system to also control AI agents operating within enterprises.