Microsoft has introduced Guardian Agents, AI agents whose job is to watch and rein in the other agents a company runs. Two security blog posts on 19-20 October call autonomous agents "agentic insiders"; Satya Nadella added a post of his own.
“We should not ask a model to be its own security authority. The controls that determine what it can access, what actions it can take, and whether it should continue operating must sit outside the model, where the model cannot bypass or tamper with them.”
Nadim Abdo, Microsoft corporate VP, Identity and Network Access, in Microsoft's security blog post, as reported by TechRadar · techradar.com ↗
Nadim Abdo, the corporate VP for Identity and Network Access, argues that agents are not hostile by design but need broad permissions to be useful, so a mistake, manipulation or compromise can do real harm. People cannot match machine speed, so separate agents supervise, adjusting an agent's permissions, behaviour and access as its risk changes.
On the practical side, Guardian Agents plug into Microsoft 365 so IT and security teams get one view of agent activity and risk. Discovery of agents running locally grows by 35 types to 60, including Claude Code, OpenAI Codex and GitHub Copilot, with finds logged in a registry of owner, activity and risk.
Microsoft's stance is that access and shutdown controls belong outside the model. Nadella framed it as separating the supply of intelligence from authority over it, so a system needs to trust the model as little as possible. It follows Project Perception and June's Microsoft cyber stack, which used agents against outside attackers.
Sources · 2