Gartner^® Names Kai a Sample Vendor in Autonomous Exposure Remediation

At Kai, we’re glad to be welcomed into the category. And not for the obvious reasons of recognition after a short time, which is nice, but mainly because this signifies the industry is evolving into a focus on innovation that’s required to address the new security remediation challenges that come with AI.  

Gartner defines Autonomous Exposure Remediation as using “...generative AI to orchestrate fully autonomous, closed-loop remediation at scale. It proactively generates, validates, and deploys security fixes across code, supply chain, and infrastructure. AER shifts security from reactive, manual patching to a preemptive discipline, neutralizing threats at machine speed while integrating with CI/CD pipelines for safe, auditable implementation.” (Emerging Tech Impact Radar: Preemptive Cybersecurity, Gartner, Elizabeth Kim and Luis Castillo, September 11, 2026)

Until now, a security leader evaluating this kind of platform had no shared vocabulary to evaluate it with. Every vendor described autonomous remediation in its own terms, which meant the claims could not be compared. A definition from Gartner helps change that. It gives buyers a standard and way to compare vendors.

What “certainty” actually requires

I argued in my last post that autonomy is downstream of certainty. Consider what has to be true before a machine closes a finding with no one watching.

The finding has to be real, not one of four duplicate identities the same vulnerability carries across three scanners. The vulnerable code path has to be reachable in this environment, not reachable in principle. The exploit has to work here, against this configuration, on this asset, given what that asset actually does. The fix has to be the right fix, and it has to not break the thing it is fixing.

Every one of those is a validation question. None of them is an automation question.  

A platform that automates remediation without answering them first has not built autonomy. It has built a faster way to be wrong. And the failure is quiet. A model handed inconsistent inputs produces a confident, fluent, plausible answer, and plausible failure is exactly the kind that survives review.

So, the order we built in was validation, then certainty, then execution. Prioritization fell out of that order as a casualty. Ranking exists because you cannot fix everything and you need a defensible way to choose. Once the false positives are gone and what remains is confirmed exploitable, choosing stops being the problem and throughput becomes the problem.

What it produced

In one deployment, Kai investigated and triaged 250 million vulnerabilities from enterprise scanners in 20 hours, 83% eliminated as benign, and auto-remediated the remainder. In another, 2.5 million software composition analysis findings across 5,000 container images were processed in under an hour, eliminating 99.5% as false positives, saving roughly 3 million engineering and security hours.  

The part the category does not yet cover

Validation and automation together do not make every fix land immediately. A patch needs a maintenance window. An owner needs time to review a change to a production system, usually ten to fifteen days. The business may not accept some changes until the following quarter. During that interval the exposure is confirmed, the fix is known, the ticket is assigned, and the asset is still exposed. Almost nobody measures that interval. Kai Watch is our answer. The moment an exposure is confirmed, Kai Watch generates and deploys context-based detection rules to the specific assets where that CVE is open, then monitors SIEM, EDR, and XDR telemetry for signs of active exploitation while the fix moves through whatever process it has to move through. It observes rather than alters, so it does not touch the asset, its configuration, or its network path, which is what makes it safe to run in production where most compensating controls are not.

It is not a substitute for the fix. It narrows the window in which the unfixed thing is also unwatched. As the category matures, I expect that interval to become something buyers ask about directly.

Every fear people hold about automated remediation, fixing the wrong thing, breaking production, acting outside its authority, traces back to acting on a finding that was never proven. Build the validation layer first and the automation becomes an ordinary engineering decision. Skip it and no amount of model capability rescues you, because the model was never the weak link.

Prioritization kept the backlog alive. Proof is what retires it.