Microsoft will add AI agents that monitor other AI agents

Microsoft will extend its security platform for AI to monitor the AI agents that operate in company networks. Hayete Gallot, an executive vice president of the company, gave the plan in a blog post. Gallot names these agents “agentic insiders”. They can get data, use software tools and make decisions with no constant human supervision. Microsoft tells that they are a risk when no attacker controls them.

How it works

Project Perception uses 3 types of AI agents:

• Red team agents find weak points.
• Blue team agents look into suspicious work.
• Green team agents repair problems.

The company wants it to find the agents in a network, give them identities and permissions, and monitor their work. It will add “Guardian Agents” that monitor other agents. They can limit permissions or stop tasks.

The proof

Microsoft gave no test results or numbers for the new work. The platform went into public preview on 3 August. Gallot tells that agents “will collaborate and collude with other agents to reach their objectives”. Gallot gave no data for this.

What is not known

Gallot tells that to monitor the work of agents is only part of the answer. Future tools must also show the cause of each step of a model. The post gives no date for Guardian Agents.

Insight: Microsoft wants controls outside the model but uses AI for them

On 10 October, Microsoft chief executive Satya Nadella wrote that controls must work apart from AI models. The new plan uses AI agents to monitor other AI agents. This reads as a race to hold the control layer for agents, and the first layer that companies use can set the rules. Leaders must check that a monitor made of AI is isolated from the model it checks.