Cybersecurity has spent years using artificial intelligence to detect suspicious behavior. CrowdStrike and NVIDIA now want AI to do something considerably more aggressive: attack a network, learn where its defenses fail, repair those weaknesses and then attack it again.
CrowdStrike has introduced SafeMind, a new family of specialized cybersecurity models and agent harnesses developed with NVIDIA technology. Instead of relying on one general-purpose AI model, the system brings together separate offensive and defensive models designed to continuously challenge each other.
The idea resembles automated red-team versus blue-team exercises — except both sides can operate at machine speed.
And if CrowdStrike’s early results translate into production environments, it could point toward a significant change in how organizations test their defenses.
SafeMind Gives AI Two Opposing Jobs
SafeMind launches around two purpose-built models.
Red Tempest plays the offensive role. It is designed for advanced attack scenarios and attempts to reproduce the behavior of AI-assisted adversaries.
Blue Solano sits on the defensive side, analyzing what happened during those attacks and creating protections intended to stop them.
But the important part isn’t simply having two cybersecurity models.
It’s what happens when they are connected.
CrowdStrike says its agent harnesses place Red Tempest and Blue Solano inside a closed feedback loop. The offensive system searches for a viable attack path. The defensive system studies the resulting telemetry, identifies gaps in coverage and generates new detections.
Then the offensive side tries again.
The process repeats, forcing both sides to adapt.
Instead of waiting for a real attacker to expose a weakness, SafeMind is designed to make AI search for that weakness first.
CrowdStrike plans to operate SafeMind natively within its Falcon platform, while trusted access to standalone models and harnesses will be offered through its Project QuiltWorks program.
NVIDIA Nemotron Provides the AI Foundation
The models weren’t created entirely from scratch.
CrowdStrike built SafeMind using NVIDIA Nemotron open models, with NVIDIA acting as an AI design partner. CoreWeave infrastructure is also being used for training and inference.
The more interesting ingredient, however, is CrowdStrike’s own cybersecurity data.
According to the company, SafeMind’s training draws from Falcon sensor telemetry alongside threat intelligence, Falcon Complete managed detection and response annotations, and knowledge accumulated through roughly 15 years of incident-response work.
That specialization matters.
A general-purpose frontier model may understand software, networks and cybersecurity concepts, but it hasn’t necessarily been optimized around the enormous volume of security telemetry generated inside enterprise environments.
SafeMind is an attempt to build models specifically for that world.
How the Attack-Defense Loop Actually Works
NVIDIA has provided a more detailed look at the experimental architecture.
The companies evaluated the system inside an isolated environment modeled on NVIDIA accelerated computing infrastructure.
The process can be simplified into four stages.
1. The offensive AI attempts an attack
A red-agent harness receives a threat-informed objective and searches for an attack path inside the controlled environment.
Its actions are recorded while CrowdStrike Falcon sensors collect telemetry.
2. The defensive AI reconstructs what happened
The defensive system receives the attack trace and corresponding telemetry.
It determines which actions can be reconstructed, which existing defenses detected the behavior and — most importantly — which parts of the attack escaped detection.
3. New detections are generated and tested
The blue-agent system creates candidate detections for the gaps it found.
Those detections aren’t simply accepted because an AI produced them.
A validation process tests them against the captured telemetry, rejects failures and sends unsuccessful candidates back for correction.
Validated detections can then be added to the defensive environment.
4. The attacker tries again
Now the red agent faces the strengthened defenses.
It attempts another route or evasion technique, producing fresh telemetry for the defensive agent.
And the cycle begins again.
NVIDIA describes the goal as continuing until the offensive system can no longer find another viable path within the modeled environment.
The Early Performance Claims Are Significant — but They Are Internal
CrowdStrike is making some striking performance claims for SafeMind.
According to the company’s own evaluations, the system produced a:
- 29% higher detection rate
- 6x faster end-to-end remediation
- 99% reduction in detection and remediation costs
NVIDIA separately describes an evaluated configuration in which the defensive Blue Solano model achieved higher accuracy at substantially lower cost than the proprietary frontier model CrowdStrike used for comparison.
Those numbers deserve attention, but they also need context.
They are company-reported evaluation results, not evidence from a broad independent benchmark across production networks.
Real enterprise environments are messy. They contain legacy software, unusual configurations, incomplete telemetry and behavior that is considerably harder to reproduce than a controlled test environment.
So SafeMind’s early results are better understood as evidence that the architecture can work — not proof that every organization deploying it will see equivalent improvements.
Why Offensive AI Is Becoming Part of Defense
There is a larger reason this technology is arriving now.
AI systems are becoming increasingly capable of performing cybersecurity tasks that once required substantial human expertise.
That helps defenders, but the same capabilities can help attackers.
An AI system capable of examining software, exploring infrastructure and reasoning through possible vulnerabilities creates a fundamentally different threat environment when it can also operate autonomously.
Traditional security tools often react to something that has already happened: suspicious behavior occurs, telemetry is collected, an alert fires and a human analyst investigates.
SafeMind proposes a different approach.
Continuously search for the attack before the real attacker does.
That doesn’t eliminate traditional detection and response. Instead, it potentially gives defenders an automated adversary that continuously pressure-tests those defenses.
This Is Not an AI Freely Hacking the Internet
There’s an important distinction here.
The architecture NVIDIA describes was evaluated in an isolated, representative environment. The offensive model is intentionally given a controlled place to behave like an attacker.
That’s very different from allowing an autonomous system to roam across arbitrary production infrastructure or public networks.
The architecture also includes validation mechanisms around the defensive side. Candidate detections are tested before being accepted, rather than allowing an AI model to blindly rewrite security policies based on its own conclusions.
Those controls become increasingly important as cybersecurity agents gain more autonomy.
An AI that merely recommends a detection rule creates one level of risk.
An AI that can attack systems, alter defenses and execute actions creates another entirely.
Cybersecurity May Become a Competition Between Machines
CrowdStrike CEO George Kurtz describes the objective as moving beyond AI that simply identifies threats toward AI capable of defeating them.
NVIDIA CEO Jensen Huang frames the same transition as a continuous competition: attackers using AI to scale offensive capabilities while defenders use AI to expand detection and response.
That may ultimately be the most important part of SafeMind.
The individual models will change. Nemotron will evolve. Competing cybersecurity vendors will build their own systems.
But the underlying architecture — an autonomous attacker continuously testing an autonomous defender — could become much more important as AI agents gain the ability to operate independently across real computing environments.
For security teams, the future may therefore look less like analysts waiting for the next alert.
It may look like machines fighting thousands of simulated battles before the real one ever begins.
Sources / Research Notes
The core technical details are based primarily on CrowdStrike’s SafeMind announcement and NVIDIA’s technical description of the attack-defense architecture. CrowdStrike’s performance figures are presented as company-reported evaluation results rather than independent benchmarks. Independent reporting was used to cross-check the broader context.
CrowdStrike — SafeMind announcement
NVIDIA — SafeMind and CrowdStrike announcement
NVIDIA Technical Blog — Adaptive Agentic Cybersecurity System
