Ignite announcements usually come with a lot of marketing, and this one does too. Below is what Microsoft announced, how the agents and playbooks work, what Microsoft's documentation says about permissions and limitations, and what security teams should check before handing work to the agents.
From July preview to Ignite GA
Microsoft's Tech Community blog for the product describes the path: "When we introduced Project Perception in July, we set out a vision for defense that could operate at the speed and scale of AI." Neowin reported the GA announcement and noted the shorter name.
Microsoft's live Ignite news coverage placed Perception inside a larger security architecture. According to that coverage, the Microsoft cyber stack has three layers: signals, sensors and actuators; context; and agents, harnesses and models. Perception sits at the top layer. It orchestrates blue, red and green team agents through playbooks, working together in a closed loop to secure your systems.
Neowin said the signals layer draws on Defender, Entra, Intune, Sentinel and Purview. It also cited Microsoft's claim of more than 165 trillion signals a day. That number comes from Microsoft's Digital Defense Report 2026, which says the company processes more than 165 trillion security signals daily. It describes Microsoft's overall telemetry, so it says nothing about how much data Perception itself handles or how well it performs.
Section summary: Perception is now generally available. Microsoft presents it as the agent layer that sits on top of its existing security products, not as a replacement for them.
Red, blue and green agents
The team names come from security exercises. Nadella summarized the roles this way: "red team agents simulate attacks, blue team agents detect attacks and green team agents fix and remediate."
Microsoft Learn's Perception overview adds more detail:
- Red team agents look for vulnerabilities before attackers can exploit them.
- Blue team agents detect, triage and investigate threats.
- Green team agents remediate problems and harden your security posture.
Microsoft says the three groups form a closed loop that keeps finding, evaluating and improving an organization's security posture.
Playbooks are how work gets assigned. A playbook gives a team of agents a set of objectives and coordinates what they do. In Microsoft Learn's words, agents share "intelligence, findings, and security context," and they surface recommendations or ask for approval when needed. Every playbook can be run manually, and some can also be triggered by automation rules. A natural-language chat can also pick the right playbook, fill in its inputs and run it.
Nadella went further on stage: "With Playbooks, you can compose teams of agents to tackle complex jobs and trigger them to take action autonomously." Autonomous here means configured autonomy. The documentation keeps approval gates in the workflow, as covered below.
What the agents do
Microsoft's application card for Perception, a responsible-AI disclosure document, lists specific workflows:
| Agent | What it does | Notable boundary |
|---|---|---|
| Triage Agent | Classifies supported alerts as malicious or false alarms and records a verdict with its reasoning | Starts automatically when a new supported alert is created; can also be run manually |
| Attack Investigation Agent | Tier-2 investigation of Defender incidents, producing a verdict, timeline, attack graph, affected entities and remediation actions | Limited to Defender incidents |
| Recon Agent | Attacker-style mapping of privilege paths, choke points and exposure risks | Read-only; Azure environments only |
| Threat Intelligence Agent | Pulls threat-actor profiles, MITRE ATT&CK mappings, IOCs, CVEs and KQL hunting queries out of threat articles | Output depends on the source material it is given |
| Posture Prioritization Agent | Ranks findings by severity, exploitability, active exploitation, internet reachability, asset criticality and attack-path context | Some Learn pages still label it preview |
The Threat Intelligence Agent has its own GA announcement. Microsoft's security blog says at Ignite, the Threat Intelligence Agent in Microsoft Perception is also becoming generally available, and that from a Threat Analytics article, security teams can now run the Protect against a threat playbook in Microsoft Perception. That playbook can chain the threat-intelligence, reconnaissance, posture-prioritization and detection-authoring agents together: read a threat report, check whether your environment is exposed, rank the gaps, and write KQL detections for them.
In the Defender portal, Perception has five sections: New Chat, Sessions, Agents, Playbooks and Insights. Some playbooks can also be launched from incident pages, alert pages and threat-intelligence articles, so analysts can start them without going to the Perception area first.
Section summary: The core is triage, investigation, threat-intel extraction, attack-path mapping and posture ranking. The Threat Intelligence Agent's GA is separately confirmed.
What's new at GA
Microsoft's Tech Community post introduces a Reduce exposure risk playbook. It calls on the Risk Reduction Agent to assess the available paths for reducing risk and recommend the right response. The same post says Perception is bringing exposure management and agents together.
Neowin reported several other additions:
- actionable plans
- new playbooks
- a preview of vulnerability shielding, which gives vulnerable resources temporary protection while teams prepare a proper patch
- more Purview integrations over the coming months
Constellation Research corroborates the shielding feature: vulnerability shielding will be available so enterprises can buy time before applying a permanent patch. Constellation also reports that Perception includes AI Readiness Initiatives in preview in January, and the cybersecurity platform will cover Microsoft, AWS and Google Cloud.
Those are announcements, and Microsoft has not yet documented them in detail. The Recon Agent's documented scope is still Azure-only. Shielding is a preview. Microsoft has not given exact timing or configuration details for each item, so budget and plan around what is documented today.
Model updates: Nadella said "We are integrating Cyber 2 with Perception," and described Cyber 2 Flash as "our most capable security model yet." The application card says Perception's agents reason with Azure OpenAI models from Foundry. Microsoft states those models are not trained on customer data from Perception sessions. Your organization's data, threat intelligence and Defender data are supplied to the model when it runs, a technique Microsoft calls grounding.
MDASH: Neowin reported that Microsoft's agentic code scanner, codenamed MDASH, is now in an expanded preview. It can analyze repositories up to 5 GB and run in regulated clouds. Microsoft's Perception blog adds that MDASH can scan larger repositories more efficiently by breaking them into logical units and analyzing them in stages. MDASH remains in preview.
How much autonomy does it really have?
Not as much as "machine speed" suggests, and that is deliberate. Microsoft's own documentation spells out the limits:
- Approval gates: For high-impact actions, agents pause and wait for a human. The application card says this design is intentional.
- Process logs: Each session shows its intermediate steps, the data it used and its reasoning. Analysts can override conclusions, stop sessions or redirect agents.
- Role-based access: Administrators decide who can run agents, who can view results and who can change settings.
- Accuracy: Microsoft warns that outputs can be inaccurate, incomplete or outdated, and tells users to validate critical results. The card lists overreliance risks such as missed threats, wrong incident conclusions, ineffective detection rules and remediation of the wrong resources.
- Data and permissions: Agents only see what their configured identities can reach. Playbooks that include an unconfigured agent can't run.
This matters because an agent that writes KQL detection rules or recommends remediation is only as good as its data access and the person checking its work. A Triage Agent that misclassifies an alert you're watching is a nuisance. If it closes the wrong one automatically, an attacker can go unnoticed.
Preview labels in the documentation
When the source research was done, some Microsoft Learn pages still said "Applies to: Project Perception," and at least one agent page still said Limited Public Preview. Expect the documentation to catch up over the coming weeks. In the meantime, don't assume every agent or feature became GA on November 17. The product is GA, the Threat Intelligence Agent is explicitly GA, and vulnerability shielding and MDASH are explicitly preview. Anything else, check its label in your tenant.
Getting started
Based on Microsoft's own deployment guidance:
- Inventory your data sources. Perception depends on connected Defender services and grounding data. Gaps in telemetry become gaps in what the agents find.
- Create agent identities with least privilege. Microsoft advises giving each agent access only to the data and actions its task requires.
- Set up RBAC for who can run sessions, view results and change settings.
- Test with realistic scenarios first. Run playbooks against representative incidents and confirm the approval gates trigger before deploying broadly.
- Add automation gradually. Start with manual playbooks, then turn on automation rules for well-understood cases such as Triage on noisy alert types.
- Review session logs regularly and use the feedback controls when agents get things wrong.
Microsoft also lists Ignite breakout sessions for the product. One is BRK318, "Project Perception: How Agentic Security Works," on Thursday, November 19. It is worth watching if you're evaluating Perception.
Analysis
Perception fits a pattern: attackers are using AI, so defenders are being sold AI. Microsoft's Digital Defense Report argues that AI is speeding up attacks and that defenders need to work at machine speed too. The argument holds up. Nobody wants a human analyst reading 3,000 alerts by hand.
There are reasons for caution, though:
- Lock-in. Perception works best on Microsoft's data. The multicloud coverage Constellation describes is a promise, while the documented Recon Agent works only in Azure today.
- Customer evidence. Microsoft named BNY, but only as exploring Perception capabilities alongside Defender, Entra and Purview. So far there are no published production results.
- Pricing and licensing were not detailed in the announcement material reviewed here. Ask your Microsoft account team before you plan a budget around it.
Perception looks like a serious attempt to have AI agents do SOC grunt work without removing human oversight. Treat it the way you'd treat a capable new tier-1 analyst: let it work, check its output, and expand its responsibilities as it proves reliable.
Update: Petri reports additional Perception agents and exposure features (November 18, 2026)
Petri reports that Microsoft has also made the Detection Authoring Agent generally available alongside the Alert Triage, Attack Investigation and Threat Intelligence agents. The original announcement material discussed detection-writing workflows, but did not clearly identify the Detection Authoring Agent as a separately GA component.
According to Petri, the Posture Prioritization Agent is generally available, contrary to the preview labels still visible on some Microsoft Learn pages when this article was researched. Admins should treat the tenant’s current feature label and licensing screen as the practical source of truth until Microsoft’s documentation is fully synchronized.
Petri also says the Recon Agent is GA, while the Web App Pen Test Agent and Risk Reduction Agent remain in preview or public preview. It further reports expanded remediation playbooks spanning Microsoft, AWS and Google Cloud environments, although Microsoft’s documented Recon scope remains Azure-only.
References
- Microsoft Perception Brings AI Agents into Security Ops - Petri IT Knowledgebase Petri IT Knowledgebase · 2026-11-18T10:46:04+00:00
- Microsoft Perception is now generally available Neowin · 2026-11-17T19:02:02+00:00
- Application card for Perception | Microsoft Learn learn.microsoft.com