Microsoft Perception is generally available. In agentic security, the playbook is the control, not the agent
Perception orchestrates red, blue and green security agents through playbooks, and now reads AWS, Okta and Salesforce signals. Its key features limit what the agents may do.

In one incident Microsoft’s security researchers describe, an attacker sent 157 push verification prompts to several employees, from about two dozen IP addresses in more than 20 countries, until enough people tapped “approve.” The stolen Okta session was then used to reach Microsoft 365. An analyst looking only at Okta logs, or only at Microsoft 365 logs, sees half of that story.
That is the case Microsoft is making for Perception, its agentic security system, which became generally available at Ignite 2026 on November 17 and gained third-party data support two days later.
What Perception is
Perception, previously Project Perception, coordinates specialized security agents in three roles. Red team agents look for weaknesses from an attacker’s point of view. Blue team agents triage alerts, investigate attacks, author detections and build threat intelligence. Green team agents prioritize weaknesses and help fix them. Playbooks combine these agents into larger jobs and can be triggered automatically or launched from Microsoft Defender.
The general availability covers the agents and playbooks first announced in July. Microsoft says the Triage Agent now covers about 80% of alerts in a typical Defender queue, and the Investigation Agent follows attacks across endpoints, identities, email, Active Directory and cloud resources.
The third-party expansion, read carefully
The November 19 post adds signals from AWS, Okta and Salesforce, ingested through Defender connectors. The scope differs by agent, and Microsoft’s own documents describe it slightly differently:
- The Investigation Agent and Detection Authoring Agent can work with any vendor data flowing into Defender, with tailored expertise for AWS, Okta and Salesforce.
- The Triage Agent is generally available for selected third-party alerts. The November 19 post names Okta and Salesforce alerts generated by Microsoft security products; the Book of News lists AWS and Salesforce. Expansion to more sources is planned.
Before planning around it, check which alert types your tenant actually receives triage for. “Supports Okta” can mean investigation over Okta data without triage of Okta alerts.
Microsoft also introduced Cyber 2 Flash, a small security model it says performs ahead of frontier models at a fraction of the cost per task, and said it is integrating Cyber 2 with Perception. For now, the Book of News lists it in preview only inside Codename MDASH, Microsoft’s code security system, and the performance claim has no independent evaluation yet.
The controls are the product
Most of the attention goes to what the agents can find. For a security leader, the more important list is the one that limits them. Microsoft’s announcement includes least-privilege access, approval controls, audit logging, usage views and a cost calculator. In the preview Reduce exposure risk playbook, the Risk Reduction Agent can coordinate patching through Intune “with IT maintaining control of deployment,” or apply temporary protection when a patch would disrupt the business.
This is the right shape. As we argued about the agentic SOC, autonomy in security is earned per action class. Triage verdicts and investigation timelines are low-risk outputs a human reviews. Disabling an account, isolating a host or pushing a patch are actions with blast radius. A playbook is where that distinction gets written down: which agent may do what, on which trigger, with whose approval.
It also matters because attackers already use agents. Google’s threat intelligence team recently logged 27 attacker actions in 23 seconds. Defense at human speed loses that race; defense at machine speed without bounded authority creates its own incidents.
Questions to settle before turning playbooks on
Which actions run without approval? List them per playbook, and start with the ones that are reversible.
What identity do the agents use? A security agent with tenant-wide rights is one of the most privileged identities you will ever create. Scope it like one, and review it like one.
How is a wrong verdict caught? Sample the Triage Agent’s closed alerts each week, especially the ones it marked benign. The 80% coverage figure says how many alerts it reads, not how many it gets right.
Where do the records go? Playbook runs should land in the same case history as human work, so an auditor can see who or what decided, and on what evidence.
What does it cost? Use the cost calculator before enabling playbooks that run on every alert. Agent-driven investigation at queue scale is a meter, not a flat fee.
Perception moves security operations toward agents that act. The organizations that benefit most will be the ones that wrote down, before go-live, where those agents must stop.