Nvidia and CrowdStrike launch SafeMind, the first agentic cyber-defense system running natively on Falcon

Jensen Huang and George Kurtz took the stage in Las Vegas to announce CrowdStrike SafeMind, the first agentic cyber-defense system that runs natively on Nvidia’s Falcon platform. CrowdStrike data show AI-based attacks jumped 89% in the past year, and the fastest eCrime breakout time now stands at 27 seconds. Human response is no longer relevant; it is documentation, not defense.

SafeMind does not rely on a static defense model. It is built around a Coevolution loop in which attacker and defender sides challenge each other sequentially, hardening the environment until attacks fail. CrowdStrike built the defense model on Nvidia’s open-source Nemotron models after post-training on 15 years of threat data and cyber-experience. Proprietary harnesses sit atop the models, turning them into autonomous agents, which Huang described as an “external skeleton” that makes the brain act.

Under the hood, two Nvidia models serve defined roles: Nemotron 3 Ultra orchestrates the defensive harness, and a tuned version of Nemotron 3 Super drives a sub-agent that writes detection rules. CrowdStrike claims the model “Blue Solano”, based on Nemotron 3 Super, achieved higher accuracy than leading frontier models at a lower cost by 99%, a figure that comes from the company’s internal estimates, not an independent benchmark.

Kurtz stressed that this is not a chat layer built on someone else’s intelligence, but a front-end model developed and trained by CrowdStrike in partnership with Nvidia. The system is delivered as a complete package, yet the models are also available separately, allowing customers to attach their own models to CrowdStrike’s harnesses—a flexibility rarely seen in products of this type. Huang added that their asymmetric advantage stems from a large community of experts who want to collaborate and keep the world safe.

Alongside SafeMind, the company unveiled Falcon IQ, intended to turn Project QuiltWorks into agentic automation of workloads, and expanded its Guardian AI safety solution. The conference attracted 10,000 security professionals from finance, health, the public sector and critical infrastructure. The message was clear: attackers had AI at the front, defenders did not, and that is changing now.