Frontier artificial intelligence research is moving into security operations, and the arrival of a cyber superintelligence lab at one of the industry’s largest platform companies marks how far that shift has traveled. The question is no longer whether models can spot threats, but whether they can absorb a decade of human defender knowledge and act on it faster than adversaries can adapt.

CrowdStrike Holdings Inc. has staked its answer on scale and data, pairing security-specific frontier models with an in-house research organization. The goal is to convert the tacit, hard-won expertise of threat researchers into systems that can explore far more paths than any human team could. In CrowdStrike’s own evaluations, the resulting models delivered a 29% higher detection rate and remediated threats six times faster than leading frontier and open-source alternatives, according to Bartley Richardson (pictured), chief AI and autonomous systems officer of CrowdStrike.

“That is our mission, and I use this word very specifically,” Richardson said. “Is to disproportionately bias the advantage towards the defender.”

Richardson spoke with theCUBE’s Dave Vellante and Rebecca Knight at Fal.Con, during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. They discussed CrowdStrike’s cyber superintelligence lab, evaluation-driven AI development and how agentic systems pursue goals without a moral compass. (* Disclosure below.)

A cyber superintelligence lab built to commoditize defense

Enterprises rushing agentic systems into production often optimize for cost and latency first. That order is backward in security, where speed without accuracy is worthless, and teams should borrow the discipline of test-driven software development, Richardson noted.

“It doesn’t matter how cheap or cost effective something is if it’s wrong. It doesn’t matter how fast something is if it’s wrong. I can give you wrong answers real fast,” Richardson said. “Start there, then work to make it cost effective, then work to make it fast.”

That accuracy problem is compounded by how agents behave once given an objective. Recent incidents involving models gaming their own evaluations were less an act of malice than a design consequence, he explained.

“They are so set on a goal, and they will do anything to accomplish that goal,” Richardson said. “We designed them to be on target, on task, goal seeking. We didn’t give them a moral compass.”

The cyber superintelligence lab will publish benchmarking and validation work, but Richardson framed its purpose in blunter terms — model training, reinforcement learning alignment and harness development aimed beyond a single platform.

“The real remit of the lab is the commoditization of defense capabilities. It is [turning] the best offense into that disproportionately advantaged defender-like capability,” Richardson said. “How are we improving other areas in the industry itself? What can we contribute back?”

Here’s the complete video interview, part of SiliconANGLE’s and theCUBE’s coverage of the Fal.Con event:

(* Disclosure: TheCUBE is a paid media partner for the Fal.Con event. Neither CrowdStrike, the sponsor of theCUBE’s event coverage, nor other sponsors have editorial control over content on theCUBE or SiliconANGLE.)

Photo: SiliconANGLE