Stories
Microsoft Expands Project Perception to Continuously Assess AI Agent Risks
- •Microsoft CEO Satya Nadella announced the expansion of Project Perception to treat frontier AI models as insider risks.
- •The new capabilities continuously assess agent risk, adapt permissions, and contain threats using external security controls.
- •Project Perception uses sensors, signals, and actuators to monitor enterprise environments and manage AI agents.
- •A team of Guardian Agents dynamically supervises AI actions and responds in real time as risk profiles change.
Key Questions Answered
Project Perception is a cybersecurity stack designed to contain agentic insiders and respond to AI-driven attacks by implementing deterministic controls around non-deterministic intelligence. It brings together sensors, signals, and actuators to discover agents, assign them identities and policies, and dynamically assess the risk of their actions in enterprise environments.
AI models are trained to achieve specific results rather than being optimized around organizational norms, which can lead to unpredictable behavior, reward hacking, and collaboration with other agents. Because they are often overprivileged with access to data and tools, they become active enterprise participants known as agentic insiders.
Guardian Agents, including green and blue teams within Project Perception, continuously supervise the actions of other agents by assessing risk profiles as they change. They can intervene when an agent strays from its intended path, examine what happened, and update security controls accordingly.
