AI agents inside companies are being recast as a new kind of insider: they need identities, rules and continuous supervision, not just one-off security checks.
Microsoft presented Perception as a system that does exactly that for enterprise agents. It brings together identity, data, cloud, endpoint, network and runtime signals, gives agents recognized identities and controlled environments, and applies policies and permissions. The announcement also introduced Guardian Agents, supervisory programs that watch agent behavior and step in when an agent’s risk profile changes.
Why this matters: companies are already giving agents broad permissions to act on their behalf. That makes them a distinct insider risk even if an attacker never compromises them. Perception treats that risk as operational, not hypothetical, so security teams can manage agents continuously instead of hoping a perimeter holds.
How it works, in plain terms: think of an AI agent as an employee who needs a badge, a job description and a manager. Perception issues the badge and job rules, watches where the employee goes and what systems they touch, and uses both fixed controls and adaptive AI to flag or stop risky actions. Guardian Agents are the on-the-floor supervisors who intervene when the situation changes.
What changes now: security and compliance teams get a concrete framework to govern agents rather than ad hoc workarounds. Deploying this will require integrating agent identity and telemetry into existing tools, and careful policy design so agents do useful work without excessive privileges.
What’s next: the real test will be whether organisations adopt these controls before incidents force them to. Will industry standards and regulators converge on a common approach to agent identity and continuous supervision? That is the question to watch.
