Microsoft proposes Guardian Agents to supervise AI agents inside companies

Microsoft says its Project Perception security system can be used to manage AI agents inside companies, with Guardian Agents supervising their actions in real time. The proposal matters because enterprise agents can have access to data, tools and critical systems, making oversight part of the product rather than a box to tick afterwards.

Microsoft AI Watch analysis

What happened

In a Microsoft blog post published on 19 October, Hayete Gallot described how the company’s Project Perception security stack could be applied to “agentic insiders”. The system would bring agents under management by assigning them identities, containers, permissions and policies, then use signals from areas including identity, data, cloud, endpoints, networks and runtime to assess the risk of proposed actions.

Microsoft says teams of “Guardian Agents” could supervise agent actions continuously and respond as risks change. Its post also argues that greater model transparency, including access to reasoning-related tokens, could help security teams understand why an agent acted. The company says transparency would add to containment, not replace it. Read Microsoft’s explanation.

Why it matters

An AI agent with access to company systems is not just a chatbot with a longer to-do list. It can take actions, and Microsoft’s argument is that those actions need monitoring and controls that adapt as the risks change. The proposed approach also puts a live debate on the table: how much visibility into a model’s internal process is useful for security, and how should that access be governed?

Our read

The useful shift here is from asking only what an agent did to asking whether its actions fit the task and permissions it was given. That is a more serious standard than trusting the agent because it works for the same organisation. Microsoft presents Project Perception as something organisations can use today, but the post is a description of the approach, not evidence of how well it performs in deployments. Security teams should look for concrete controls, audit trails and clear limits on what Guardian Agents can access.

What to watch

• Which Project Perception controls Microsoft makes available for enterprise agents.
• How the system explains, records and responds to risky agent actions.
• What safeguards govern access to model reasoning-related tokens.

Discussion spark: Should companies let security systems inspect a model’s reasoning-related tokens to supervise agents, or should oversight rely on actions and audit trails instead?

not affiliated with or endorsed by Microsoft