NVIDIA and CrowdStrike test agentic red-blue loop
NVIDIA and CrowdStrike ran offensive and defensive agents against each other in an isolated replica of NVIDIA's accelerated computing infrastructure, using open Nemotron models to write and validate detection rules without human handoffs.
In a technical blog post, NVIDIA describes a joint evaluation with CrowdStrike of a closed-loop attack-defense agent system. A red-agent harness executes a threat-informed attack path inside a representative environment while CrowdStrike Falcon endpoint sensors capture telemetry; a blue-agent harness reconstructs the event sequence, identifies visibility and detection gaps, and generates candidate detections. A validation harness backtests each candidate against the captured telemetry and returns failures for correction. An independently seeded attack then retests the same objective, and the red agent adapts until no viable path remains.
The defensive configuration paired NVIDIA Nemotron 3 Ultra for orchestration with a fine-tuned Nemotron 3 Super acting as a bounded expert for detection authoring, running inside CrowdStrike SafeMind. Supporting mechanisms included a schema knowledge base to prevent invented Falcon fields, telemetry grounding, and automated artifact linting.
CrowdStrike says its Blue Solano defensive model is 13% more accurate than the leading proprietary frontier model tested at 97% lower cost — a figure from internal evaluations, not independent audit. The target environment was built from a sanitized natural-language specification of NVIDIA infrastructure and reviewed by NVIDIA security staff for realism.
The post frames the work as a directional case study rather than a general benchmark, covering one scenario family with small detection sets. It also cites the Open Secure AI Alliance as a vehicle for expanding open models and harnesses for security work.