Microsoft blitzed Ignite 2026 with agents that hunt, agents that pen-test, agents that fix and agents that police other agents.
Microsoft Chairman and CEO Satya Nadella promised defenders “machine speed.” Charles Lamanna, executive vice president of Copilot, Agents and Platform, said customers are already “managing over 50 million agents” in Agent 365. Aleš Holeček, president and chief architect of Microsoft Security, declared that “AI has changed the physics of cyber.
Time to pop the hype bubble.
SPB held Microsoft’s Ignite security pitch up against Microsoft’s own earlier announcements to sort what’s new from what’s warmed over. Keeping it short, because it’s Friday, here’s what matters to AI security pros, identity practitioners, AppSec teams and the SOC and MSSP crowd.
The biggest Ignite headlines weren’t about security. Microsoft made Word, Excel and PowerPoint generally available inside Copilot and pitched Microsoft IQ, a layer that feeds agents company data through hundreds of connectors, as Nadella argued that “Copilot is the new OS for work.” That’s big news for IT and business buyers, though Office Watch counted only 76 of 215 new features as generally available today. SPB is sticking to the security pitch.
Ignite isn’t Microsoft’s security show, per se, or its developer show. It’s the company’s flagship enterprise IT and cloud conference for the people who buy, deploy and run its products – IT pros, admins, business leaders and partners. That said, this year’s Ignite puts “Security & Compliance” and “AI security” on the agenda beside Copilot and cloud.
AI Security Pros: The Agents Need a Babysitter
The most useful news Ignite got little stage time. Defender’s runtime protection for local AI agents is generally available, and it can audit or block Claude Code, GitHub Copilot CLI, OpenClaw and ChatGPT Desktop on Windows endpoints. An Agent 365 demo showed Microsoft now discovers 65 types of local agents on employee devices and gives each a “dynamic agent risk” score built from runtime behavior, exposure and potential blast radius.
Holeček explained why that matters. Agents are “not inherently malicious,” he said, but “they are frequently overprivileged, have access to a broad set of tools and high-value data, and are increasingly more autonomous, executing business processes around the clock.” He called them “agentic insiders.” Microsoft’s own Perception launch post concedes agents “can make mistakes, be manipulated, or be compromised.”
The controls to rein them in are mostly still on the way. Conditional Access, Privileged Identity Management and an MCP firewall for agents are still in preview in Entra, where Microsoft says tool and agent abuse already accounts for 12% of the attack activity it sees on Azure AI workloads. Execution Containers, which a demo used to “scope the agent’s access,” debuted as an early preview at Build in June. And Agent 365 itself is a year old, as Lamanna noted: “One year ago here at Ignite we announced Agent 365” (Microsoft, Nov. 18, 2025).
Then came the credibility problem. Nadella told the arena that “what a model tells us about itself is not even independent evidence, and therefore cannot be trusted,” and that “one model monitoring another model has the same issue.”
The result, he said, is an “unacceptable nested set of black boxes.” He announced a Microsoft AI Transparency Initiative, with “already 40 organizations” signed on, to push for standards that let customers “verify its actions independently.”
Not long after, Mustafa Suleyman, CEO of Microsoft AI, said Microsoft’s new MAI Cyber 2 Flash model is “ahead of GPT-6 Astra on CyberGym end-to-end at less than 2% of the cost per task.”
Translation: Microsoft’s compact in-house security model, built for its own products and successor to August’s MAI Cyber 1 Flash, edged OpenAI’s frontier model on CyberGym-E2E, a UC Berkeley test that makes AI agents find, prove and patch real open-source bugs. The score was 92.3% to 91.5%, after Microsoft dropped 92 of the benchmark’s tasks. That’s Microsoft’s number about Microsoft’s model.
Identity Practitioners: Passkeys are Old News: The SMS Bill Isn’t
Microsoft is still touting passkeys as “now the default for new accounts” in Entra. It announced that July 13 and began rolling it out Sept. 1. The same post tackles the harder news that Microsoft stops providing SMS and voice authentication in Entra on Feb. 1, 2027. After that, holdouts buy it from a telecom partner through the Microsoft Security Store and pay for it themselves. With Microsoft’s own Entra team saying 57% of organizations still use phishable sign-in methods, plenty of tenants will feel it.
What’s genuinely new is agent identity lifecycle management, now generally available.
Every agent in Agent 365 gets “a unique identity in Entra, so admins can govern it like any user,” said Jeremy Legg, AT&T’s CTO of AT&T Services, during a demo during the day-one keynote.
Microsoft’s Entra blog says a default policy disables agent identities idle for 60 days. On stage, the demo flagged agents unused for 90 days, “a threshold admins can adjust.” AT&T’s Jeremy Legg, whose company runs more than 60,000 agents, summed up the risk: “Forgotten agents can become rogue agents.”
He likened them to “digital employees” that keep access to “your privileges, your access layers, your authentications.”
AppSec Teams: ‘Coordinated Disclosure is Collapsing’
The day’s most quotable security moment wasn’t a product pitch.
“Just this month’s Patch Tuesday, we addressed more than 1,000 vulnerabilities. Compare this to the fewer than 100 in March. That is 10x increase in less than a year,” Holeček said, adding that “the same pattern holds true for all major platforms, browsers, and open source packages.”
Microsoft’s Defender team separately puts October’s count at 1,413 CVEs, “the largest Patch Tuesday on record.”
The economics have flipped, too.
“Bounty programs pay about $300,000 for high-quality vulnerabilities,” Holeček said, and AI-assisted researchers, “and also the bad guys,” are “now finding them for less than $5.” Meanwhile, “the time window between discovery and exploitation is approaching zero,” in part “because the process of responsible and coordinated disclosure is collapsing.”
That’s a big admission from the vendor that ships the patches. It’s also the setup for the upsell. Defender’s new vulnerability shielding, in preview, applies “targeted, temporary protection” between disclosure and patch.
In the demo, David Weston, Microsoft’s corporate vice president of enterprise and OS security, introduced as “the architect of Perception,” faced a Patch Tuesday of “about 1,500 distinct CVEs” and roughly 52,000 unpatched devices. He patched clients right away and used “our new endpoint shielding technology” to buy time on servers. Microsoft pitched “predictive shielding” a year ago at Ignite 2025.
Perception’s red-team agents will draw AppSec interest. Nadella said the web app pen-test agent takes on “the auth, the APIs, injection attacks, business logic.” Microsoft’s own benchmark scores it at 31.0%, against 15.2% for Codex, and says it stayed within its testing boundaries in 91% of safety-constrained runs. Read the other way, it strayed out of bounds about once in every 11.
SOC Teams and MSSPs: GA, But You’ve Seen Most of It
Perception going generally available was the headline. The product isn’t new.
Microsoft Perception is Microsoft’s “agentic security system”. It consists of teams of AI agents that hunt threats, probe your defenses and fix what they find, fed by Defender, Entra, Purview and Sentinel. Its general availability was the security headline, but the product isn’t new. Microsoft unveiled it July 27 and opened a public preview Aug. 3.
Nadella described red agents that “simulate attacks,” blue agents that “detect attacks” and green agents that “fix and remediate.” Holeček said it rides on “more than 500 security connectors” and “more than 165 trillion signals every day” from 1.6 million security customers. Microsoft says the triage agent can clear about 80% of a Defender alert queue, a figure it labels “not a guaranteed triage rate.”
Nadella’s showcase was older still.
In his keynote, he described an unnamed independent researcher who turned an autonomous AI attack tool loose on a network protected by Microsoft Defender. “The agent kept trying to move laterally across the network,” Nadella said, “but Defender automated attack disruptions really kept containing those compromised accounts and isolating devices.”
Automatic attack disruption was introduced at Ignite in October 2022.
One customer voice came from BNY CIO Leigh-Ann Russell, a Perception preview customer. “When the bad guys have AI at their fingertips and go in AI time, our teams have to also have great AI tooling and also work in AI time, not human time,” she said. The tools, she added, have to keep “our cyber teams really in control of the defenses.”
The late-but-real item: security tools can now run outside the Windows kernel, generally available on Windows 11. Microsoft promised that in November 2024, four months after the CrowdStrike crash.
The SIEM Gets Cheaper. The Bill Doesn’t
The license line matters more than the demos. Microsoft is folding Sentinel’s SIEM into Defender for Defender Suite, E5 and E7 customers. Case management, playbooks and workbooks are generally available; the full SIEM is still in preview. Logs from Microsoft’s own security tools come in free. Everything else costs $2.40 per GB, against $4.30 for pay-as-you-go Sentinel in East US, by one MSSP engineer’s math. And managing Sentinel in the Azure portal ends March 31, 2027.
“Microsoft just priced the SIEM like a feature,” CyberMerge wrote the day after the keynote. That engineer, Bartosz Wysocki, head of security engineering at a UK MSSP, called it “a licence change wearing an agentic costume,” though “what’s underneath is actually good.” For MSSPs like his, “‘we run your SIEM’ gets harder to sell.” Yih Khai Wong, senior research manager at IDC Asia/Pacific, told CSO that consolidating “creates greater dependency on that vendor’s roadmap, pricing model, availability, and security resilience.”
The bill climbs from there. Defender Suite lists at $12 per user per month on top of E3; E5 runs $60 and E7 $99. Premium governance of non-Microsoft agents requires E7 or Agent 365, $15 per user on its own, which Gartner called “a work in progress with limited net new functionality,” The Register reported in March. Multi-tenant agent management for MSSPs is in public preview. Perception has no published price, though Weston’s demo showed a “FinOps dashboard” tracking what each agent run costs.
One name never came up on stage: Security Copilot. A year ago, Microsoft promised it “will be included for all Microsoft 365 E5 customers.” It isn’t in the Perception launch posts either.
One More Thing
Nadella opened not with agents but with the Nov. 12 outage that knocked out Outlook, Teams and Copilot.
“Last week, in some sense, was really a reminder of how critical digital infrastructure is as the dial tone for the world,” he said. “I know we have work to do to provide that resilience that you can count on.”
Microsoft has made versions of that promise before, with the Secure Future Initiative in November 2023 and Nadella’s “security above all else” memo in May 2024.
Brad Smith, Microsoft’s vice chair and president, speaking on Microsoft’s Ignite livestream after the keynote, framed the bigger picture. AI is “increasingly misused” by “a growing set of cyber threat actors,” he said, and “the barriers to entry have fallen.”
His conclusion: “Now law and regulation need to catch up,” and Microsoft’s answer is “giving our customers more tools to manage AI.” Conveniently, that’s what Microsoft spent the week selling.
Top photo: Microsoft Chairman and CEO Satya Nadella speaks at Microsoft Ignite 2026 in San Francisco. Photo by Dan DeLong, courtesy of Microsoft.