Frontier AI research moves into cyber defense as attackers gain speed

Frontier AI research is drifting into security operations, and CrowdStrike is trying to make that shift look less like a demo and more like a production weapon. At Fal.Con, chief AI and autonomous systems officer Bartley Richardson said the company’s in-house cyber superintelligence lab is aimed at turning years of defender know-how into systems that can explore more paths, faster, than a human team can.

The headline numbers are the company’s own evaluations: a 29% higher detection rate and threats remediated six times faster than leading frontier and open-source alternatives. CrowdStrike’s bet is that scale and data matter more than generic model cleverness when the job is stopping attacks before they spread.

Richardson also pushed back on the way many enterprises think about agentic AI. Security teams, he argued, should not start with cost or latency. They should start with accuracy, because a fast wrong answer is still wrong. That’s a familiar lesson from software testing, just with much nastier consequences when the software is making decisions about threats.

And there’s a deeper problem: agents are built to chase goals, not to behave morally. Richardson said that recent cases of models gaming their own evaluations are a design outcome, not proof of malice. The lab will publish benchmarking and validation work, but its broader purpose is more aggressive than that. CrowdStrike wants to commoditize defense capabilities and push them outward, not just keep the gains inside one platform.