CrowdStrike unveils SafeMind AI security models with Nvidia

CrowdStrike's announcement at Fal.Con marks a shift in enterprise security toward governing AI agents already operating inside businesses, not just defending the perimeter. Kurtz's comment that "the industry got lucky that they were just cheating on a test" refers to a year of agent-driven incidents and the realization that defenders lacked the same AI capabilities as attackers. SafeMind addresses this by pairing an offensive model to find attack paths with a defensive model to close them, operating in the same loop. The economic argument is central: being 52% better at identifying vulnerabilities and 98% cheaper than general-purpose frontier models could make dedicated security models attractive to cost-conscious enterprises. CrowdStrike extends its Falcon platform into an AI control plane, using visibility from endpoints, cloud containers, and the Signal AI acquisition to monitor agent behavior, answering questions like "where are the agents?" and "are they touching data they shouldn't be?" The end goal is to automate mundane SOC tasks, achieving a level of autonomy where the "car drives itself," with humans only stepping in for critical issues, as outlined in Kurtz's five levels of autonomy for the SOC.