Security agents need a reviewable operating model

The announcement

On March 24, 2025, Microsoft introduced Security Copilot agents addressing areas including phishing, data security, and identity management. An announcement is not proof of availability in an individual tenant, and current licensing and product documentation should guide implementation.

Supporting reference: Microsoft unveils Microsoft Security Copilot agents and new protections for AI ↗

Design a bounded experiment

Start with a reviewable task and a representative set of cases. Include ambiguous cases rather than only obvious examples, and decide in advance when an analyst must take over. Keep the source evidence available to the reviewer so that a convincing explanation is never mistaken for a verified conclusion.

Evaluate mistakes explicitly

For a phishing review pilot, distinguish missed threats from unnecessary escalations. Track the time spent correcting the output and whether the analyst can understand why a decision was proposed. A single productivity number hides the differences that matter to the incident response team.

Make accountability visible

Document who owns the pilot, what actions it may perform, and how access can be withdrawn. An agent should fit into an operating procedure that someone can inspect, change, and stop when the evidence does not support continued use.