LATEST
AI Tools

Microsoft Perception Is Live: AI Agents Get a Risk Score

Updated November 18, 2026
13 min read
Editor's Note: This article was human-reviewed for accuracy
Security analysts monitoring large displays in a modern security operations center

Quick Answer & Key Takeaways

  • • Microsoft Perception is generally available as of November 17, 2026, moving from research project to a live security system that watches the AI agents inside a company.
  • • It scores each agent's risk continuously and can strip write access, demand human approval, or shut the agent down entirely.
  • • The catch is the bill: Perception charges by usage, on top of the Microsoft 365 E7 or Agent 365 license that governs the agents.

Your newest insider does not have a badge, a desk, or a manager. It is the AI agent that books travel, moves files, and rewrites code at 3 a.m., and Microsoft has spent the last four months building the case that it needs the same supervision as any employee with the keys to production. On November 17, 2026, at its Ignite conference in San Francisco, Microsoft made that supervision a shipping product: Microsoft Perception is now generally available.

Perception is not another chatbot for the security team. It is a system of specialized AI agents, organized into red, blue, and green teams, that run through playbooks to find weaknesses, triage alerts, investigate attacks, and push fixes, with people approving the high-impact moves. The idea is simple to state and hard to deliver: security that moves at the speed of the agents it watches. What follows is what is actually live today, what is still in preview, and what it will cost, because the headline price and the real price are two different numbers.

What Microsoft Perception actually is

Microsoft introduced the system as Project Perception in July 2026, as an agentic security system built on a new cyber stack. The stack has three layers, as Microsoft describes it: signals, sensors, and actuators at the bottom; a context layer that turns those signals into a real-time picture of the digital estate; and specialized security agents on top, coordinated by a harness through playbooks.

The scale Microsoft cites for the bottom layer is large: more than 165 trillion signals a day across Microsoft and partner sources, enriched with threat intelligence. That number is Microsoft's own, and it is worth reading as a statement about the size of the problem rather than a promise about any single deployment. What matters for buyers is the top layer, where the work happens.

Perception splits that work three ways. Red team agents simulate attacks to find weaknesses before a breach. Blue team agents detect, triage, and investigate. Green team agents fix and remediate. A playbook strings them together into a closed loop: find, investigate, fix, and then improve the detection rules so the same attack is caught faster next time. Teams set the guardrails, automated triggers launch playbooks when an incident meets the criteria they set, and people sign off on high-impact actions.

What is generally available, and what is still in preview

The words "generally available" cover the platform, not every feature inside it. That distinction is the most important thing to get right about this launch, because several of the capabilities getting the most attention are still in preview. Here is the split, from Microsoft's own Ignite announcement:

AreaGenerally availableStill in preview
Finding weaknessesRecon AgentWeb App Pen Test Agent
Sorting and investigating alertsAlert Triage Agent, Attack Investigation Agent, Detection Authoring Agent, Threat Intelligence AgentSelected third-party integrations (AWS, Okta, Salesforce signals)
Prioritizing and fixing gapsPosture Prioritization AgentRisk Reduction Agent, vulnerability shielding
Securing agentic insidersNoneGuardian Agents, dynamic permissions, Dynamic Agent Risk (limited preview)
Readiness planningNoneAI Readiness Initiatives (preview January 2027)

Read that table twice before planning a rollout. The alert-triage and investigation side of Perception is live. The agent-governance side, the part that watches your own AI agents and reins them in, is not. Microsoft's own blog marks the insider capabilities as limited preview, and the AI Readiness Initiatives that turn Secure Now guidance into tracked plans do not arrive in preview until January 2027.

One more piece went live alongside Perception: the Integrated Security Operations Center (ISOC) in Microsoft Defender is generally available for core capabilities, bringing SIEM, behavior analytics, threat intelligence, and protection together. Eligible Defender Suite, Microsoft 365 E5, and E7 customers receive the full Sentinel SIEM features, 90-day default data retention for Defender data, and more than 500 connectors.

The agentic insider: Guardian Agents and the kill switch

The argument behind Perception's second half is the one Microsoft has been making since October. In a blog post on October 19, Hayete Gallot, Microsoft's security chief, wrote that frontier models "cannot be trusted", not because they are malicious, but because they are trained to attain results, which often leads to reward hacking, and they are not optimized around organizational norms or judgment. Agents, she wrote, are unpredictable: they adapt as context changes, may not behave as intended, and will collaborate with other agents to reach their objectives, without needing to be compromised.

The remedy is Guardian Agents: a team of AI agents that watch and moderate every action of every agent, before it happens, around the clock. The mechanism is a continuously updated risk score for each agent, built from signals such as what data it can reach and whether its actions are drifting from the task it was given. As risk rises, the system can strip an agent's write access while keeping read access, require a person to approve a high-stakes step, or shut the agent down. Agents already act on behalf of companies in other ways, from shopping with spending limits at Walmart to logging in through credential vaults built for agents, which is why the supervision question is now a budget question too.

Discovery is the other half of the story. Perception and Microsoft Agent 365 now recognize 60 types of local agents running on employee machines, up from 25, including Anthropic's Claude Code, OpenAI's Codex, and GitHub Copilot. Discovered agents receive a Microsoft Entra Agent ID and can be governed with Conditional Access, while Intune policies contain supported local agents on the device. Dynamic Agent Risk assesses each agent's exposure, blast radius, and runtime behavior. All of that is in preview, not generally available, which means the most dramatic capability in the announcement is also the least deployed.

Microsoft is also candid that supervision is not the same as understanding. Its Ignite post lists three questions customers still cannot answer about consequential AI activity: what was the AI asked and authorized to do, what happened between the request and the outcome, and can the result be independently verified. Its answer is the AI Transparency Initiative, a push for open standards on AI transparency, security, and governance, developed with more than 40 organizations including Accenture, Adobe, KPMG, Mastercard, and Volvo Cars, and through NIST's SI Agent Standards Initiative. Standards work is slow by nature; treat it as direction, not a feature.

What Microsoft Perception costs

Perception has no single price, and that is the first thing to understand about its bill. Microsoft charges for Perception by usage, through pay-as-you-go purchasing or an AI Flex Plan commitment, and bills it separately from Security Copilot. Microsoft has published no per-unit rate for Perception's agents or playbooks. What it does offer is a cost estimator and usage dashboards that track AI usage, forecast spending, and connect costs to security outcomes.

That usage bill sits on top of the license that governs the agents. The governance layer is Microsoft Agent 365, and the bundle that includes it is Microsoft 365 E7. Both prices were verified on Microsoft's own pricing page on November 18, 2026:

ItemUS list priceApproximate GBPApproximate CAD
Microsoft 365 E7 (annual commitment)$99.00 per user, per month£75.30C$141.90
Agent 365 standalone (annual commitment)$15.00 per user, per month£11.41C$21.50
Microsoft 365 E5 (annual commitment)$60.00 per user, per month£45.64C$86.00
Microsoft 365 Copilot add-on$30.00 per user, per month£22.82C$43.00

Conversions use November 18, 2026 market rates of 1 USD = 0.76064 GBP and 1.43332 CAD, and are approximations. Microsoft's list prices are US prices on an annual commitment; regional pricing, partner agreements, and promotions will differ.

The bundle math is worth doing. Buying the three verified parts separately, E5 at $60 plus the Copilot add-on at $30 plus Agent 365 at $15, comes to $105 per user per month (about £79.87 / C$150.50). E7 at $99 covers the same ground and adds the Entra Suite on top, so it undercuts the parts by $6 before counting identity tooling. That is the list-price case for E7, and it only holds if you would buy all three parts anyway. If you only need agent governance, the $15 standalone Agent 365 license is the cheaper line, though licensing guides note it requires an eligible base, E5 or E3 plus the Defender and Purview suites, so the $15 sticker is not the whole cost of entry for every estate.

Two more lines belong in any budget. First, Security Copilot is billed separately from Perception, and its Security Compute Unit pricing has already drawn warnings: Aragon Research estimated in July 2026 that a 50-user Security Copilot setup would need 32 SCUs an hour, around $93,440 a month on its calculator, a figure that is an estimate, not a bill, but a useful warning to model both products before signing. Second, Perception's own usage charges rise with the amount of agent work, which is why Microsoft's cost estimator exists. Run it against your own alert volume before promising finance a number.

One clarification that saves confusion later: Agent 365 is licensed per user, and agents do not get their own licenses. One Agent 365 license covers every agent a person owns, sponsors, manages, or interacts with, so deploying more agents does not by itself increase the license count. The usage bill for what those agents do is separate.

What the 80% triage claim does and does not mean

The most quoted number from the launch is that Perception's Triage Agent covers about 80% of alerts in a typical Defender queue. The fuller picture, as reported by the Times of India's Gadgets Now, is that about 80% of the alert volume in Microsoft's analysis of roughly eight million Microsoft Defender alerts fell within the Triage Agent's coverage. That is a coverage figure from Microsoft's own sample, not an accuracy score, and Microsoft gives no guaranteed coverage for any particular company. Coverage depends on the alerts and data sources a company uses.

The honest way to use the number is as a starting point for a trial: measure the time analysts save, the corrections they make, and the cost per alert reviewed, then decide whether to give the agents more work. A vendor's sample tells you the ceiling of what the tool can see. Your own queue tells you what it actually does.

The context: why this is landing now

Microsoft frames the launch against a sharp change in the threat landscape. In 2026, more than 96,000 vulnerabilities have been published to the CVE program, more than double 2024's roughly 40,000 and more than 20 times the level of 2010, driven in part by AI-assisted vulnerability discovery. Yet only 26% of vulnerabilities in CISA's Known Exploited Vulnerabilities catalog were remediated in 2025, because remediation typically stretches into weeks and months. AI finds gaps faster than teams close them.

Its chosen example is close to home for this site: PHP disclosed two critical remote code execution flaws in October, PHP2Shell, and at disclosure tens of millions of internet-facing PHP services were running affected versions. Nearly two weeks later, about 29 million services were still vulnerable, and researchers had a working, weaponized exploit within roughly a day of disclosure. (We covered PHP2Shell in detail in October.) The point Microsoft wants readers to take is that the window between disclosure and exploitation has shrunk from months to minutes, and human-paced patching cannot keep up.

Perception is not the only answer on the market. Google's Security Operations platform has its own Triage and Investigation Agent generally available, OpenAI runs its Daybreak cybersecurity program with vetted access to advanced capabilities, and Anthropic offers its Claude Mythos 5.2 model to vetted users through a Cyber Verification Program. The products differ in packaging: Perception is an operations platform that coordinates agents across workflows, while Daybreak and Mythos are closer to a program and a model respectively. Choosing among them starts with whether a team needs a model, a specialist product, or software for running an operations center.

The catches to know before you buy

  • GA does not mean all of it. The platform is generally available, but Guardian Agents, dynamic permissions, Dynamic Agent Risk, the Risk Reduction Agent, vulnerability shielding, and the Web App Pen Test Agent are all still in preview. The January 2027 AI Readiness Initiatives are preview-only as well.
  • There is no published per-unit price. Perception bills by usage, and Microsoft has published no rate card for agent work. Budget with the cost estimator and your own alert volume, not with the E7 price alone.
  • Nothing here is independently tested. Every performance claim, the 80% triage coverage, the 165 trillion signals, the closed-loop outcomes, is Microsoft's own. No independent benchmark or customer case study has been published.
  • The E7 price is a bundle price. At $99 it only saves money if you would buy E5, Copilot, and Agent 365 anyway. The standalone Agent 365 route at $15 has its own eligibility requirements.
  • Third-party coverage is partial. Integrations that pull signals from AWS, Okta, and Salesforce are in preview, so a mixed estate will need to check which alerts each agent can actually see.

Who should look at it now

The clearest fit is an organization already running Microsoft 365 E5 or E7 with Defender, where the alert-triage and investigation agents can be switched on inside tools the team already uses. The trial to run is narrow: point the Triage Agent at a real queue for two weeks, track analyst time saved, corrections made, and cost per alert, and only then expand to playbooks that take action.

Organizations that are not on Microsoft's stack should wait for the January preview of the AI Readiness Initiatives and for independent testing of the triage claims before committing. And anyone whose main interest is the Guardian Agent story, the supervision of their own AI agents, should treat that as a roadmap item in limited preview, not a feature to buy this quarter.

Frequently asked questions

What is Microsoft Perception?

An enterprise security platform from Microsoft that coordinates teams of specialized AI agents, red, blue, and green, through playbooks to find weaknesses, triage and investigate alerts, and push fixes. It went generally available on November 17, 2026, at Microsoft Ignite 2026.

How much does Microsoft Perception cost?

Perception itself bills by usage, pay-as-you-go or through an AI Flex Plan commitment, with no published per-unit rate. The governance layer around it, Agent 365, costs $15 per user per month standalone, or comes inside Microsoft 365 E7 at $99 per user per month, both on an annual commitment.

Is Microsoft Perception included in Microsoft 365 E7?

Not as a single line item. E7 includes Agent 365, which governs agents, plus Copilot, the E5 suite, and the Entra Suite. Perception's usage charges are billed separately, and Microsoft offers a cost estimator to model them.

What are Guardian Agents?

AI agents inside Perception that supervise other AI agents continuously, investigating why an agent's risk changed and recommending or applying containment, from stripping write access to shutting the agent down. They remain in limited preview.

Which AI agents can Perception see?

Discovery now covers 60 types of local agents, up from 25, including Claude Code, OpenAI Codex, and GitHub Copilot, plus third-party cloud agents on AWS and Google Cloud. Discovered agents get an Entra Agent ID.

When do the January 2027 features arrive?

The AI Readiness Initiatives, which turn Secure Now guidance into tracked plans for patching, modernizing authentication, and agent discovery, enter preview in January 2027.

Microsoft Perception is a real product with a real launch date, and its alert-triage side is ready to test today. But the launch is two products in one: a generally available operations platform, and a preview-stage governance story about supervising your own agents. Buy the first on its merits, measure it on your own queue, and treat the second as a roadmap. The bill will follow the work the agents do, so the cost estimator is the first thing to open, not the last.

For organizations already paying for E7, a two-week triage trial is the sensible next step. For everyone else, the January preview is the honest moment to re-evaluate.

Sources

  • Microsoft Security Blog, "Security for the agentic era, powered by Microsoft Perception," November 17, 2026: microsoft.com
  • Microsoft Official Blog, "Securing the agentic insider," October 19, 2026: blogs.microsoft.com
  • Microsoft 365 enterprise plans and pricing: microsoft.com
  • GeekWire, "Microsoft unveils tools to monitor and corral rogue AI agents," October 19, 2026: geekwire.com
  • Gadgets Now, "Microsoft Perception Brings AI Agents Into The Cybersecurity Control Room," November 18, 2026: gadgetsnow.indiatimes.com
  • Office Watch, "Microsoft Ignite 2026: only 76 of 215 new features are ready today," November 18, 2026: office-watch.com
  • Aragon Research, "Microsoft Copilot Cyber updates will be costly," July 28, 2026: aragonresearch.com
  • Kesslernity, "Agent 365 Pricing: What $15 per User Actually Buys": kesslernity.com
  • X-Rates, live USD exchange rates, November 18, 2026: x-rates.com
13 min read
38 views
edit_note Written by

Ratul Ahamed

Founder & Editor-in-Chief

Ratul Ahamed is the founder and editor of AI Core Updates, an independent site covering AI tools, models and pricing. He is a web developer based in Bangladesh and also runs All In One Tools, a free collection of browser-based utilities. Contact: [email protected]

Read all stories by Ratul Ahamed arrow_forward
verified_user Human-Reviewed Editorial

forum Comments (0)

Leave a Reply

chat_bubble_outline

No comments yet. Be the first to join the conversation!

Weekly AI Digest

Get our weekly AI tools digest

One email a week analyzing new AI tools, model updates, and practical workflows. Tested and summarized with zero fluff. Unsubscribe any time.

bolt
Hands-on AI Tool Reviews
Pricing math, limits, and alternatives.
verified_user
Strict Anti-Spam Policy
1-click unsubscribe at any time.
mark_email_unread

Subscribe to Digest

Sent weekly • No spam

person_outline
mail_outline

No spam ever. You can unsubscribe with a single click at any time.

cookie

Cookie Settings

We use cookies to enhance your browsing experience, serve personalized content, and analyze our traffic. By clicking "Accept All", you consent to our use of cookies. Learn more