CrowdStrike has introduced CrowdStrike SafeMind, an agentic AI system purpose-built for cyber defense[1][2]. It pits an offensive model and a defensive model against each other inside the same loop, then turns the outcome into working detections. The foundation is NVIDIA's open Nemotron models, and both companies' CEOs took the stage together at Fal.Con 2026 in Las Vegas on September 1[1].
Attackers Had Frontier AI. Defenders Did Not.
Roughly 10,000 security professionals attended Fal.Con 2026[1]. George Kurtz, CEO and founder of CrowdStrike, framed the problem there: attackers had frontier AI and defenders did not, and that was the real gap[1].
The company put numbers behind it. AI-enabled attacks rose 89 percent over the past year, and the fastest eCrime breakout time, the shortest interval between initial intrusion and lateral movement, has fallen to 27 seconds[1]. At human speed, CrowdStrike argues, a response is no longer defense. It is documentation.
Jensen Huang, founder and CEO of NVIDIA, made the same point from the other side: attacks are now automated, so defense has to be automated too[1].
Two Models, Red and Blue, in One Loop
SafeMind is a family of security-specific models and harnesses developed by CrowdStrike's Cyber Superintelligence Lab[2]. A harness is the outer machinery that turns a model into an agent that actually operates. Huang described it as the exoskeleton around the large language model, which serves as the brain[1].
Two models launch with the system[2].
Red Tempest is the offensive release, built for advanced attack scenarios and designed to emulate AI adversaries. Blue Solano is the defensive release, protecting enterprise assets by deploying the battle-tested measures defenders use in the field. Running the two against each other continuously, in a coevolution loop where each side hardens the other, is the core of SafeMind[1][2].
The training data is spelled out as well. It draws on Falcon sensor telemetry, CrowdStrike threat intelligence, Falcon Complete MDR event annotations, and fifteen years of incident response fieldwork[2].
On the technical side, NVIDIA Nemotron 3 Ultra orchestrates the defensive agent harness, while a fine-tuned Nemotron 3 Super powers the rule-generation sub-agent[1]. CrowdStrike's internal evaluations found that Blue Solano, based on Nemotron 3 Super, delivered higher accuracy than leading frontier models at 99 percent lower cost[1]. Measured as a whole system against leading frontier models and open-source baselines, the company reports a 29 percent higher detection rate, 6x faster end-to-end remediation, and 99 percent cost savings on detection and remediation[2].
A closed frontier model cannot be post-trained on a company's own threat data without shipping that data outside. CrowdStrike cites exactly this in explaining why it started from open Nemotron[1]. In security, being able to inspect what is doing the defending is a practical requirement.
SafeMind can run as a complete system, but the models can also be used on their own, and customers can pair their own models with CrowdStrike's harnesses[1][2]. Separating the models from the operating layer is a deliberate design choice.
A Digital Twin of NVIDIA's Own Network
The test environment was disclosed too. NVIDIA said it evaluated the SafeMind models and harnesses in a high-fidelity cyber agent environment running as a simulation of the NVIDIA network[1].
That environment is a digital twin of NVIDIA's own accelerated computing infrastructure, validated against NVIDIA's real threat landscape[1]. Inside it, a red-team agent runs Recon, Assault and Compromise sub-agents to execute attack paths, while a blue-team agent monitors through Falcon sensors, generates detection candidates, validates them and promotes them. What comes out is usable as detections that block attacks[1].
Huang noted that this basic framework, an adversarial model working against a digital twin with a defender model in a continuous cat-and-mouse loop until the environment learns to secure itself, applies to robotics, edge computing and enterprise computing as well[1].
Falcon IQ and a Workforce of 50 Agents
At the same event, CrowdStrike announced Falcon IQ[1][2]. It operationalizes the Project QuiltWorks partner program through agentic workload automation[1].
Falcon IQ runs more than 50 agents as a single unified agentic workforce, automating the most time-intensive workflows in assessment, prioritization and remediation[1]. It runs on Charlotte AI AgentWorks, CrowdStrike's no-code agent development platform, whose agentic engine is also powered by Nemotron models. Partners use Falcon IQ to deliver customized findings, recommendations and executive outputs to their customers[1]. CrowdStrike also expanded its Guardian AI safety solution[1].
Training and inference run on CoreWeave's AI Cloud[2]. Both companies emphasize the vertical stack, from the chips up through the models to the harnesses that act on what those models find.
Summary
What sets SafeMind apart is that the product is the system itself, with offense and defense living side by side, rather than a frontier model used as a detection tool. The headline figures, 29 percent higher detection, 6x faster remediation and 99 percent cost savings, all come from CrowdStrike's own evaluations, and third-party verification has yet to arrive[2]. Even so, choosing an open model that can be post-trained on proprietary threat data, and separating models from harnesses so customers can mix in other models, gives enterprises a concrete option in how they buy security AI.
Source: https://blogs.nvidia.com/blog/nvidia-crowdstrike-fal-con-2026/
