CrowdStrike Unveils SafeMind, an Agentic Security System Built on NVIDIA Nemotron

CrowdStrike unveiled SafeMind, an agentic cybersecurity system developed in collaboration with NVIDIA, at Fal.Con 2026 in Las Vegas. Developed by the CrowdStrike Cyber Superintelligence Lab, the system stems from the necessity that defense must be automated to counter automated attacks.

How SafeMind Works

SafeMind combines CrowdStrike's custom agentic harness with NVIDIA Nemotron-based defense models to form a continuous coevolution loop where attacks and defenses challenge and improve each other. This process is designed to continuously strengthen security in customer environments, ensuring attacks fail.

• Model Composition: The defense models were built by post-training NVIDIA Nemotron open models with CrowdStrike's cyber experience and threat data.
• Performance: The Blue Solano model, based on Nemotron 3 Super, achieved higher accuracy than major frontier models while reducing costs by 99%.
• Flexibility: The system can be operated as a whole or used with models independently, and customers can combine their own models with CrowdStrike's custom harness.

Technical Advantages and Strategy

NVIDIA's Nemotron is a free open model provided to create domain-specific AI, allowing security teams to post-train with their own threat data without sending data externally. This approach is impossible with closed frontier models and plays a crucial role in ensuring transparency for the defending entity.

Jensen Huang likened the harness to an 'exoskeleton' for LLMs, explaining that it enables LLMs to function as agents optimized for specific domains. CrowdStrike operationalized Project QuiltWorks as agentic workload automation via Falcon IQ and expanded its Guardian AI safety solutions.