CrowdStrike and Nvidia put attacking and defending AI agents in a continuous test loop

CrowdStrike has introduced SafeMind, an agentic security system that repeatedly pits an attacking AI against a defending one. The offensive agent searches for an exploit in a simulated environment; the defensive agent generates and validates detections, then promotes protections intended to block the attack.

The system uses Nvidia’s Nemotron open models with CrowdStrike’s threat data and security-specific agent software. Nemotron 3 Ultra coordinates the defensive workflow, while a fine-tuned Nemotron 3 Super powers a rule-generation sub-agent. CrowdStrike says customers can also pair their own models with its agent framework, rather than relying only on the bundled models.

SafeMind will be built into the CrowdStrike Falcon platform. Nvidia is testing the models and agent framework in a digital twin of its own computing infrastructure, where red-team agents run reconnaissance, attack and compromise tasks while blue-team agents monitor Falcon sensors and prepare defenses. The companies’ performance and cost claims come from internal evaluations, so the practical measure will be how reliably the loop finds real weaknesses without creating unsafe automated actions.