AIOps for Security in 2026: Vendors Moved, Gates Didn’t

AIOps for Security in 2026: Vendors Moved, Gates Didn’t

Two product announcements arrived within a week of each other this year, and together they marked a turning point that slipped past most headlines. One assistant would begin closing suspicious mail without waiting for a human first. Another would quietly retire a correlation engine thousands of teams had built their workflows around. Both are part of the same story: 2026 is the year the AIOps for Security market — a loose, still-unstandardised label for security work assisted by machine learning and LLMs — stopped asking permission before it acted. This piece is a tour of what actually moved, told in plain language, with the evidence in the background doing its usual quiet work of separating progress from theatre. I have kept the tone of a colleague walking you through a year of release notes at the end of a long week, because the release notes themselves are dense and the stakes are not. The conclusion, argued below, is simple: the vendors moved fast, but the gates that decide what a system may safely do have not moved at all — and that gap is where buyers live now.

What Changed for AIOps for Security Buyers in 2026

The Microsoft headline came first. The company’s Phishing Triage Agent closes mail on its own from the moment it is switched on; hunt through the documentation as of October 2026 and you will find no observe-only mode to start with. Contrast that with Google’s approach: its Triage and Investigation Agent holds to level one — it recommends and drafts, a person disposes — and clears each triage in about a minute, which is genuinely fast without being brave. Meanwhile, inside Microsoft’s own estate, the Fusion correlation engine in Sentinel is being stood down, replaced by Defender XDR correlation, and the deadline for every workspace is 31 March 2027. Three different philosophies shipped in one year: act immediately, recommend first, and migrate or else. Watching which philosophy other vendors copy over the next four quarters will tell you more about this market than any benchmark slide, because copying is how this industry signals what it believes buyers will tolerate. There is also a quiet asymmetry worth noticing: the acting-first agent chose its own observe stage by skipping it, while the recommending agent makes the human the observe stage by design. Both are defensible products. Only one of them asks your team for permission first, and that difference should price into the decision more heavily than either accuracy figure.

The Year the Analysts Refused to Cheer

While product launches multiplied, the measurement community answered with numbers. ITBench, presented at ICML 2025, found autonomous agents completing just over a quarter of its compliance-and-security scenarios — 25.2%, to be precise. CyberSOCEval, a 2025 collaboration between Meta and CrowdStrike, ranged from 23 to 34 percent on malware analysis and 43 to 53 percent on reasoning over threat intelligence; blind guessing manages 0.63 and 1.7 percent on those tasks. Gartner, reading the same room, parked the AI-SOC-agent category on this year’s inflated-expectations peak and warned about AI washing — claims polished well past their proofs. The pattern deserves a name of its own: capability advancing in single-digit increments while deployment confidence advances in double digits. Nothing about the technology forbids this; nothing about the technology excuses it either. There is history here, too. When Sommer and Paxson surveyed machine learning for intrusion detection back in 2010, they observed that the field’s techniques were rarely employed in real operations precisely because errors cost so much — a caution that has outlived every hype cycle since. Flight operations readers will recognise the shape — a gap between what the aircraft can do and what the crew believes it can do is called exactly one thing.

The Migration Clock Nobody Escaped

The Fusion-to-XDR transition merits separate mention — it is the least glamorous and most consequential item of the year. Teams built playbooks, dashboards, and muscle memory around Fusion’s correlations; those now carry an expiry date of 31 March 2027. Ask any analyst what lives inside those rules and you will hear institutional memory: which product versions misbehave after patching, which login sequences precede account takeovers, which alerts are pure noise on Monday mornings. Rebuilding that into a new correlation layer is not a checkbox, it is an archaeology project with a deadline. Migration deadlines concentrate the mind in a particular way: they force the correlation logic into the open, where someone must decide what to keep, what to rebuild, and — this is the opening — whether an AI-assisted layer should sit where the old engine used to. The honest answer is that a correlation replacement is a detection-engineering project, not an AI upgrade, and the two should not be blurred. The lesson generalises across every announcement this year: dates attached to architecture are real; dates attached to intelligence are marketing until measured. Treat the first kind with respect and the second with a lab.

Where AIOps for Security Defaults Meet Everyday Risk

Beneath the launches, quieter settings shifted the risk surface. Security Copilot enables customer-data sharing by default unless the organisation carries E5 or E7 licensing — meaning a procurement decision silently becomes a data-governance decision. Gemini’s Security Ops processing runs on global endpoints, though prompts stay out of model training — which answers the training question but not the location question. Agents now read whatever arrives — including text an attacker wrote — and the UK’s NCSC said plainly in December 2025 that this exposure cannot be filtered away: the model is an inherently confusable deputy, the weakness may never vanish, and deterministic controls plus least privilege must carry the defence. Add the small print from multi-scanner uploads, where submitted files can reach other customers, and 2026’s real security story comes into focus: the products are more capable and more porous at once. A captain once told me the hardest part of new avionics is remembering which switches default to on; fleets learned to write that list down, review it after every software load, and brief it before the first flight. Security teams should do the same, switch by switch, this quarter — and then again after every renewal, because defaults migrate too.

Reading 2027 From Here

Forecasting from 2026’s evidence, three trajectories seem durable:

  • Act-first deployments keep shipping, so buyers — not vendors — must supply the observe stage. The autonomy ladder’s price list makes that practical: recommend-only needs roughly two hundred graded alerts with thirty-plus genuine ones; closing needs a miss bound at or under 2% proven across 150-plus genuine cases, with decoys and a blind holdout; reversible moves need errors under 1% across 300-plus action-worthy verdicts, with rehearsed rollback.
  • Evaluation professionalises, because the benchmarks made average performance public and nobody can un-publish it. Expect per-category reporting, seeded canaries at fifty per quarter with withdrawal after three closures, and drift monitors to become standard contract items rather than exceptional asks.
  • The 2026 numbers — a quarter of scenarios resolved, half of reasoning tasks, and precision that dazzles while misses decide — age into the field’s baseline, the way early automation rates did in every industry before this one.

The full record behind each claim above is collected in the Ops Atlas briefing covering AIOps for Security, which reads as a dossier of this exact transition, and updates as the market moves.

There is a final, human note worth keeping. Private aviation learned to love the fast-moving market — anyone who has studied how empty leg jet flights actually work knows inventory shifts hourly and the deal is only real when the terms are written down. AIOps for Security is developing the same texture: capabilities appear fast, prices of trust move daily, and the buyer who thrives is the one who reads the manifest before boarding. The travellers’ version of that discipline lives in this traveller’s walkthrough of empty-leg flying — verify the operator, confirm the route, know the exit. Apply it to every agent demo this year and 2027 arrives as a plan rather than a surprise.