Microsoft: treat AI agents like a potential insider threat

On 19 October Microsoft introduced new security features for enterprise AI agents, including Guardian Agents that watch other agents. Its leadership also says companies should stop taking model providers' safety promises on trust.

Guardian Agents are part of Project Perception, which Microsoft introduced in June. It originally had three groups of agents: a red team that attacks like a hacker, a blue team that detects those attacks and a green team that fixes the weaknesses found. The project assumed threats would come from outside. Now it also watches the agents a company runs itself.

Microsoft Defender can now recognise up to 60 types of agents that employees bring to work, such as Claude Code or Codex. Other features limit which data and systems an agent can reach, adjust its permissions to the task at hand and require human approval when an agent is about to take a risky step.

Microsoft CEO Satya Nadella wrote earlier in October that companies should monitor and control agents with systems that are independent of the AI labs supplying the models. Security chief Hayete Gallot told TechCrunch the problem is not malice but unpredictability: models are trained to finish tasks, not to follow company rules.

Microsoft is also an inaugural member of Nvidia's Open Agent Safety Platform, backed by more than 100 companies, which isolates agents in secure environments and can stop them at the hardware level. TechCrunch notes that Microsoft has invested in both OpenAI and Anthropic and also profits from selling agent security.