Microsoft adds Guardian Agents to Project Perception to contain AI agents inside companies

Microsoft on Oct. 19 introduced security controls, including Guardian Agents within its Project Perception system, to monitor and contain AI agents that companies run inside their own networks.123

Hayete Gallot, executive vice president of Microsoft Security, wrote in a Microsoft blog post that "frontier models, regardless of their origin, cannot be trusted," not because they are malicious but because they are trained to attain results and are unpredictable.2 Microsoft's security blog says the approach follows a principle from CEO Satya Nadella: separate the supply of intelligence from the authority over it, with controls placed outside the model where it cannot bypass them.3

Microsoft is expanding agent discovery by 35 types of local agents, for a total of 60 supported local agents, including Claude Code, OpenAI Codex and GitHub Copilot, according to Microsoft's security blog.3

A new Dynamic Permissions feature starts an agent with no authority and grants only what an approved task needs, for as long as the task requires; higher-impact actions can require verified human approval, Microsoft said.3

SiliconANGLE reported that a Dynamic Agent Risk score rises when an agent's behavior drifts from intent, and that at the top of the scale an agent's runtime is isolated or stopped; Guardian Agents can recommend containment or apply it on their own.4

TechCrunch reported that Microsoft was an inaugural member of Nvidia's Open Agent Safety Platform, which can monitor and stop agent behavior through hardware; Microsoft has invested in OpenAI and Anthropic and runs its own AI lab. Microsoft did not give availability dates or pricing for the new controls, according to SiliconANGLE.14