NVIDIA and CrowdStrike Unveil SafeMind Agentic Security System

At Fal.Con 2026, NVIDIA and CrowdStrike announced a partnership launching CrowdStrike SafeMind, an agentic cybersecurity system that simulates controlled attacks, converts Falcon telemetry into detection rules, and back-tests detection against new attack paths. In replayed-attack tests, average detection of unknown attacks improved from 16.5% to 41.9% — a landmark case of AI-agent red/blue teaming in security.

Confirmed

• SafeMind comprises two models built on NVIDIA Nemotron open models: red-team model Red Tempest for finding attack paths and blue-team model Blue Solano for patching vulnerabilities, running on the CrowdStrike Falcon platform.
• The architecture is validation-first: the defensive side combines a pattern knowledge base, Falcon telemetry and domain context with specialized models, tools and validation mechanisms; the offensive side simulates controlled attacks to generate new attack paths and back-test detections, with the two models continuously adversarial in a closed loop.
• In tests on replayed attacks, the jointly optimized open-source pipeline raised average detection from 16.5% to 41.9%.
• Per Gene Kim, the partnership was announced by Jensen Huang at the CrowdStrike event; the two "frontier-level" models were trained on CrowdStrike data, and CrowdStrike built a "digital twin" environment for NVIDIA.

Why it matters

• Detecting unknown attacks (zero-days, novel TTPs) is a chronic SOC pain point; the 16.5%-to-41.9% gain demonstrates the potential of agent-vs-agent adversarial loops with telemetry feedback in real back-tests.
• The deal is a flagship case of NVIDIA landing its Nemotron open-model stack in enterprise security and may push more vendors toward agentic architectures.

2026-09-02 ~ 2026-09-03 · 8 related posts

• [source] NVIDIA and CrowdStrike test AI agents to defend against unseen cyber attacks — NVIDIAAI · 2026-09-02
• NVIDIA and CrowdStrike build adaptive agentic cybersecurity system — NVIDIAAI · 2026-09-02
• NVIDIA AI agents boost unseen attack detection rate to 41.9% — NVIDIAAI · 2026-09-02
• [source] CrowdStrike launches SafeMind with offensive and defensive AI models — Codeblix_Ltd · 2026-09-02
• NVIDIA and CrowdStrike partner on frontier-level red and blue team AI models — RealGeneKim · 2026-09-02
• CrowdStrike and Nvidia launch SafeMind, offense-defense model loop at Fal.Con — eyishazyer · 2026-09-03

2 near-duplicate retellings: nordicinst · nvidia