17:37Machine-translated from Azerbaijani by AIRead in Azerbaijani
Microsoft Unveils Guardian Agents System for Monitoring AI Agents
Microsoft on Monday unveiled new security tools for keeping AI agents used by companies under control. The main innovation is the Guardian Agents system. The system monitors autonomous agents as potential threats that could emerge from within a company. Microsoft security chief Hayete Gallot said that advanced models cannot be fully trusted, regardless of which company developed them. This is because even if these models are not malicious, they may behave unpredictably while carrying out a task.
An AI agent is a program that works independently with data and digital tools to achieve a given goal. Because such systems are granted extensive permissions, any mistake can cause serious harm. Microsoft CEO Satya Nadella emphasized that agents can access sensitive data and make important decisions. Therefore, the control system that determines what they can do must be separate from both the model and the company that developed it. Project Perception was presented in June as an AI-powered cybersecurity system. Initially designed to counter external attacks, the system used 3 groups of agents. Red agents searched for vulnerabilities, while blue agents investigated the threats they found. Green agents resolved problems and made the necessary fixes. Guardian Agents applies the same protection to agents operating within a company. Because agents make decisions within seconds, monitoring is continuous and faster than humans can keep up with. Project Perception and Microsoft Agent 365 share threat alerts, risk data and a common registry. This allows IT and security teams to monitor all agents from the same dashboard. The defense system covers 7 areas: identity, permissions, data, devices, network, cloud and the environment in which the agent operates. The system first determines who the agent belongs to, why it was created and what it does. It then restricts the agent’s access to data and tools and protects the resources it uses. As the agent’s behavior changes, the security rules are also updated.
Microsoft’s agent identification system will cover 35 additional types of native agents, including Claude Code, OpenAI Codex and GitHub Copilot. Microsoft Defender can detect up to 60 types of agents that employees bring into their work environments. OpenClaw is among them. Microsoft Entra Agent ID is getting new identity types for agents operating within companies that were previously unidentified. This allows security teams to see who each agent belongs to. Dynamic Permissions grants an agent only the access permissions needed for a specific task. Dynamic Agent Risk looks for dangerous behavior. When risk rises, Dynamic Conditional Access can block the operation, request human approval or disconnect the agent from the system. Microsoft is one of the first members of the Nvidia Open Agent Safety Platform initiative, which is supported by more than 100 companies. The platform keeps agents in a protected environment. On Surface Laptop Ultra devices, a similar control system has been connected to Nvidia chips.
An AI agent is a program that works independently with data and digital tools to achieve a given goal. Because such systems are granted extensive permissions, any mistake can cause serious harm. Microsoft CEO Satya Nadella emphasized that agents can access sensitive data and make important decisions. Therefore, the control system that determines what they can do must be separate from both the model and the company that developed it. Project Perception was presented in June as an AI-powered cybersecurity system. Initially designed to counter external attacks, the system used 3 groups of agents. Red agents searched for vulnerabilities, while blue agents investigated the threats they found. Green agents resolved problems and made the necessary fixes. Guardian Agents applies the same protection to agents operating within a company. Because agents make decisions within seconds, monitoring is continuous and faster than humans can keep up with. Project Perception and Microsoft Agent 365 share threat alerts, risk data and a common registry. This allows IT and security teams to monitor all agents from the same dashboard. The defense system covers 7 areas: identity, permissions, data, devices, network, cloud and the environment in which the agent operates. The system first determines who the agent belongs to, why it was created and what it does. It then restricts the agent’s access to data and tools and protects the resources it uses. As the agent’s behavior changes, the security rules are also updated. Microsoft’s agent identification system will cover 35 additional types of native agents, including Claude Code, OpenAI Codex and GitHub Copilot. Microsoft Defender can detect up to 60 types of agents that employees bring into their work environments. OpenClaw is among them. Microsoft Entra Agent ID is getting new identity types for agents operating within companies that were previously unidentified. This allows security teams to see who each agent belongs to. Dynamic Permissions grants an agent only the access permissions needed for a specific task. Dynamic Agent Risk looks for dangerous behavior. When risk rises, Dynamic Conditional Access can block the operation, request human approval or disconnect the agent from the system. Microsoft is one of the first members of the Nvidia Open Agent Safety Platform initiative, which is supported by more than 100 companies. The platform keeps agents in a protected environment. On Surface Laptop Ultra devices, a similar control system has been connected to Nvidia chips.