Microsoft adds just-in-time permissions and risk controls for AI agents

Microsoft has announced new controls for AI agents in Microsoft 365 E7 and Project Perception, including task-limited permissions and access that can tighten as an agent’s risk rises. The practical shift is towards giving agents only the access they need, then adjusting or stopping them when warning signs appear.

Microsoft AI Watch analysis

What happened

The new Dynamic Permissions feature starts agents with no access of their own and grants only what they need for a particular task, for as long as that task lasts, according to Microsoft. Dynamic Agent Risk continuously assesses risk using signals from identities, data, devices, cloud, networks and the agent’s actions.

Microsoft also says Dynamic Conditional Access can respond to rising risk in stages: an agent might lose permission to change files while keeping read access, lose access to some data, need human approval or be stopped. The company says it can now detect 35 additional types of agents on staff devices, bringing the total to 60. The Next Web reports that Microsoft has not said when the new features will be available or whether they will cost extra.

Why it matters

As agents gain access to workplace data and tools, permissions set once and left alone can be a poor fit for actions that change from task to task. Microsoft’s approach aims to make access temporary and responsive, rather than treating an agent’s initial authorisation as a permanent all-clear.

The controls are a product announcement, not evidence that they have been released or that they will prevent every harmful action. Availability, pricing and how the risk scoring performs in real deployments remain open questions.

Our read

The useful idea here is not that another layer of monitoring will magically make agents safe. It is that an agent should have to earn access for each job, and that access should be reducible when circumstances change. That is a better starting point than handing over the keys and hoping the dashboard looks reassuring.

What to watch

• When the controls become available and whether they cost extra.
• Which agent types the expanded detection can identify in practice.
• How Microsoft explains the signals behind risk changes and the actions administrators can take.

Discussion spark: Should workplace AI agents start with no access and request permissions for each task, or does that create too much friction for useful automation?

not affiliated with or endorsed by Microsoft