Microsoft Perception and the Agentic SOC: What Ignite 2026 Actually Delivered
Every Ignite has a theme you can’t escape in the hallways. This year in San Francisco, it’s the “agentic SOC”. It’s been hard to get away from, and session BRK324 was literally titled “The Alert Is Dead”.
I’m naturally sceptical of anything that claims to make the alert queue disappear. But Microsoft has shipped real things this week, so let’s look at what’s actually been announced, what you can use today, and the questions I’d be asking before betting a SOC on it.
Perception: From Project to Product
Project Perception, introduced in July, is now Microsoft Perception. Microsoft’s Security blog and the Perception blog both describe it as generally available (see the note on the date below).
The model is three teams of specialised agents, orchestrated by playbooks:
- Red team agents find and validate potential paths to compromise from an attacker’s perspective.
- Blue team agents triage alerts, investigate attacks and turn threat intelligence into detections.
- Green team agents prioritise exposures and coordinate remediation.
You can start a playbook directly, describe what you need in chat, or have playbooks trigger automatically when an incident meets criteria you set. Microsoft is clear that your team sets the guardrails and signs off on high-impact actions.
A Note on the GA Date
Microsoft’s 17 November announcements say “now generally available”, although the Perception blog’s intro says it “will be”, and the speakers in partner session BRK494 gave 19 November. Either way, it’s GA this week. Check your tenant before you plan around a specific date.
What’s New in the Agents
The highlights from the Perception blog:
- Triage Agent now covers about 80% of alerts in a typical Defender queue. Read the footnote: that figure comes from Microsoft’s analysis of around 8 million alerts, and it’s aggregate coverage, “not a guaranteed triage rate for an individual customer”.
- Attack Investigation Agent reasons across endpoints, identities, email, Active Directory and cloud resources, rebuilds the attack timeline, and recommends prioritised response actions.
- Third-party signals (preview, per the Security blog): Attack Investigation can use Okta and AWS, Detection Authoring can use Okta, Salesforce and AWS, and Triage can use Okta and Salesforce. The Book of News lists Triage as AWS and Salesforce instead, so check which applies in your tenant. More vendors are promised.
- Threat Intelligence Agent turns external threat articles, including other vendors’ research and government advisories, into structured findings. The Detect emerging threats playbook then has the Detection Authoring Agent produce tenant-aware KQL. That detection is tested against historical tenant data to estimate alert volume before an analyst approves it. That’s the bit I like most: it’s tuning before deployment, not after the flood.
- Red team: the Run web pen test and Protect web app playbooks, using the Web App Pen Test Agent (preview), test an authorised web app and can recommend targeted fixes, such as WAF rule changes in Azure Front Door.
On the green side, the Security blog says the Posture Prioritization Agent is GA and the Risk Reduction Agent is in public preview. Agent-security playbooks for “agentic insiders” remain in limited preview (the Book of News calls it private preview).
What’s GA and What’s Preview
The statuses are scattered across several announcements, so here’s my summary of what Microsoft has published:
| Capability | Status | Source conflict? |
|---|---|---|
| Microsoft Perception (platform) | GA | 17 vs 19 Nov |
| Posture Prioritization Agent | GA | |
| Risk Reduction Agent | Public preview | |
| Third-party signals (Okta, AWS, Salesforce) | Preview | Triage vendors differ |
| Web App Pen Test Agent | Preview | |
| Agentic insider playbooks | Limited preview | Book of News says private preview |
| Core ISOC in Defender | GA | |
| Full SIEM breadth and 90-day Defender retention in ISOC | Preview | |
| Vulnerability shielding | Preview, on by default |
ISOC: The Foundation Underneath
In Microsoft’s own description of its cyber stack, Perception “brings the top of the stack together” and ISOC is “part of that foundation”. For day-to-day operations, it’s arguably the more important of the two. The Defender & Sentinel Ignite blog confirms that core Integrated Security Operations Center (ISOC) capabilities in Defender are now generally available. That covers case management, automation playbooks (not to be confused with Perception’s agent playbooks), workbooks, UEBA, workspace creation and third-party ingestion.
Still in preview: the full breadth of SIEM capabilities for eligible Defender customers, 90 days of built-in Defender data retention, SOC optimisation recommendations, Content Hub, detections as code and automation audit logs. Existing Sentinel customers can opt into ISOC.
Vulnerability Shielding: Buying Time Before the Patch
Vulnerability shielding is in preview and, per Microsoft, on by default in Defender. It applies targeted, temporary controls to vulnerable systems in the gap between exposure and remediation. The Defender & Sentinel Ignite blog puts the scale of the problem like this: “In October 2026 alone, 1,413 CVEs were disclosed, marking it the largest Patch Tuesday on record.”
It works in two modes, and the difference matters (vulnerability shielding blog):
- Proactively, the Risk Reduction Agent in Perception can recommend a shielding plan through the Reduce exposure risk playbook when you can’t patch straight away. Microsoft says “nothing is implemented without your approval”. In the preview this covers server-side remote code execution vulnerabilities.
- During an active attack, Defender’s attack disruption applies shielding controls autonomously. This is Defender behaviour, not a Perception playbook, so the approval guardrails I described above don’t apply in the same way.
The controls themselves are things like reducing network reachability, restricting access paths and tightening permissions. Shields are time-bound, retire automatically once the patch lands, and can be rolled back in one click, with exclusions managed through Intune.
In my view, this is the most quietly significant announcement of the week for anyone running a patching service. It doesn’t replace patching, and Microsoft says so too, but it changes the conversation about the window of exposure. Because the attack-disruption side is on by default, I’d want to understand exactly what it can change, and set exclusions for anything that must never be touched, before a customer asks why their server suddenly stopped accepting a connection it used to.
What Will It Cost?
Perception is priced separately from Security Copilot, as the BRK494 speakers pointed out. Microsoft’s Perception product page lists pay as you go, or an up-front commitment through an AI Flex Plan or MACC. There are new usage dashboards and a cost calculator, and the Insights page in Perception includes a Cost and Usage dashboard.
That’s sensible, but in my experience consumption-based security tooling has a habit of surprising people. Agents working while nobody’s watching is the whole point, and also the whole budget risk. As a BRK494 speaker put it, AI is non-deterministic, and that makes it hard to put a fixed price on every playbook. I haven’t seen a hard spending cap for Perception documented, so watch the Cost and Usage dashboard from day one and set budget alerts on whatever billing scope it lands in.
How I’d Evaluate It
- Pick one playbook with a measurable outcome. Triage is the obvious one. Take a sample of your own alerts and measure analyst time saved, how often analysts overturn the verdict, and cost per alert.
- Run it in parallel, not instead. Let the agent triage while your analysts still do, then compare.
- Agree a budget and set up cost alerts before you enable automatic triggers. Use the cost calculator to estimate it first.
- Try Detect emerging threats against a recent advisory and review the KQL it produces. It’s a good test of whether the agent understands your tenant.
- Read up on vulnerability shielding in your environment. It’s on by default in preview, so know what it’s doing, and set your exclusions.
The MSP View, and the Questions I Still Have
For managed security providers, this changes the economics of L1 triage. If an agent can reliably reach a first verdict on a large share of your queue (and that’s what your pilot should prove), the value we add moves towards investigation, response, tuning and governance. That’s a good thing, but it means re-thinking how managed SOC services are scoped and priced.
My open questions:
- Multi-tenant. I haven’t seen Microsoft explain how Perception and ISOC work across many customer tenants, the way an MSSP actually operates. That’s the first thing I’ll be asking the product team.
- GBP pricing and purchasing route for UK customers and partners.
- Which agents and playbooks are GA versus preview in any given tenant. The announcements mix both.
The alert isn’t dead. But for the first time, I think a good chunk of the volume might become someone else’s problem, and that someone is an agent. Just make sure a human is still checking its homework.