Breaking

NVIDIA and CrowdStrike launch SafeMind, Nemotron-based agentic cyber defense, at Fal.Con 2026

September 1, 2026 at 17:54 EDT

Foundation Models
AI Agents
Security

NVIDIA and CrowdStrike on September 1, 2026 introduced CrowdStrike SafeMind, a family of purpose-built security models and agent harnesses built on NVIDIA Nemotron open models, aiming to automate red-team and blue-team defense at machine speed. The companies unveiled the system at CrowdStrike’s Fal.Con 2026 in Las Vegas, where NVIDIA founder and CEO Jensen Huang joined CrowdStrike founder and CEO George Kurtz before a crowd of about 10,000 security professionals.

September 1, 2026 · NVIDIA · CrowdStrike

SafeMind: red vs blue at machine speed

CrowdStrike launched SafeMind at Fal.Con 2026, pairing NVIDIA Nemotron open models with offensive Red Tempest and defensive Blue Solano harnesses that coevolve on Falcon telemetry.

CrowdStrike internal evaluations vs frontier / open baselines

Faster end-to-end remediation

29%

Higher detection rate

99%

Cost savings on detect and remediate

NVIDIA evaluation: mean backtest detection

Optimized open pipeline vs default Nemotron 3 Ultra harness (same recorded-attack backtest).

16.5%

Ultra + default harness

41.9%

Optimized open pipeline

2.5x mean detection vs the default harness — a system-level gain, not an isolated model ablation.

Closed loop

Red Tempest attacks twin → Blue Solano writes detections → Validate and retest

Nemotron 3 Ultra orchestrates; fine-tuned Nemotron 3 Super authors rules. Native in Falcon; standalone via Project QuiltWorks.

What held up

Live-fire: 45% of open detections generalized vs 29% frontier; three open “gold” detections covered all eight unseen attacks.

Limits noted

One scenario family, small detection sets, limited benign traffic, and harness failures in three of eight live-fire runs. Directional case study, not a general benchmark.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.