The development
What happened
Microsoft outlined on 19 October how Project Perception can contain enterprise agents it calls “agentic insiders.” The system, first announced in July, combines permissions and runtime signals with continuous supervision of attempted actions. Microsoft presents this as a response to risks from internally deployed agents, including those that have not been compromised.
Key facts
Signals span identity, data, cloud, endpoints, networks and runtime; managed agents can receive identities, containers, policies and permissions.
Dynamic context assesses the risk of attempted actions, while Guardian Agents supervise activity and respond as risk changes.
Security practitioners review traces and apply judgment to adjust the surrounding controls.
The post gives no independent effectiveness benchmark or deployment-coverage figures. Greater visibility into model reasoning is a proposed next step, rather than a demonstrated ability to read intent.
The context
Why it matters
This approach puts enforcement around the agent’s tools and execution environment as well as its model. Its value depends on which actions the controls can intercept, how permissions are enforced and whether intervention works under realistic failures.