NVIDIA and CrowdStrike created an agentic system that automates the red-team/blue-team feedback loop.
Nemotron 3 Ultra handled orchestration across attack reconstruction, planning, and tool use, while a post-trained Nemotron 3 Super generated and repaired detection rules. In the evaluation, the specialized open-model pipeline was the only approach to produce rules that passed every quality gate, catching all eight fresh attacks.
The project highlights a practical pattern for domain-specific agents: pair a reasoning model with a specialized, post-trained model and validate every output against real tools, telemetry, and deterministic checks.
Resources:
👉 Learn more about the offensive - defensive agentic system: Building an Adaptive Agentic Cybersecurity System with NVIDIA Nemotron | NVIDIA Technical Blog
📖 Try the [NVIDIA Nemotron 3 open models](Nemotron AI Models | NVIDIA Developer)
🧑💻 Explore [NeMo Megatron Bridge](NeMo Megatron Bridge — Megatron Bridge)
🧪 Evaluate specialized agents with [NeMo Gym](NeMo Gym | NeMo Gym)
⚙️ Post-train models with [NeMo RL](NeMo RL Documentation — NeMo-RL)