CrowdStrike introduces SafeMind, a new AI system for cybersecurity in which an offensive and a defensive model are constantly pitted against each other. The models are based on Nvidia Nemotron and are integrated into the Falcon platform.
At launch, SafeMind consists of two specialized models. Red Tempest acts as a red team and attempts to identify attack vectors and vulnerabilities. Blue Solano represents the defensive side and must then take measures to eliminate the identified risks.
What’s particularly interesting is the way CrowdStrike combines these two models. They operate within a closed cycle in which Red Tempest attempts an attack and Blue Solano responds. The results are then used for the next round. CrowdStrike describes this as a system in which the models continue to evolve together.
Proprietary security data for training
For development, CrowdStrike uses Nvidia’s open Nemotron models as a foundation. The company then adds its own security data and expertise. The training data comes from sources including telemetry from Falcon sensors, threat intelligence, annotations from Falcon Complete MDR, and fifteen years of incident response. According to CrowdStrike, the telemetry covers endpoints, identity systems, cloud workloads, and data stores, among other things, and encompasses trillions of events per day.
Nvidia is participating in the project as an AI design partner. CoreWeave also provides infrastructure for training and running the models.
With this approach, CrowdStrike aims to differentiate itself from the use of general-purpose large AI models for security tasks. Instead of a single model assigned various tasks, separate models are being developed for offensive and defensive tasks. This distinction is becoming more relevant as general-purpose state-of-the-art models are proving increasingly capable of performing security tasks. Recent incidents, including the attack on Hugging Face in which AI agents played a role, have also demonstrated that such capabilities can be used offensively.
So-called “harnesses” play an important role in this. This software layer determines how the models are controlled and what actions they can perform. The harnesses are not limited to CrowdStrike’s own models; other frontier and open-source models can also be used with them.
From analysis to action
With this approach, CrowdStrike aims to take AI beyond simply detecting or analyzing security issues. Within permitted parameters, the models must also be able to take actual actions to mitigate identified risks. This aligns with the broader trend in which security vendors are deploying AI agents for an increasing number of aspects of investigation and incident response.
SafeMind will become a direct component of the Falcon platform. Separate models and harnesses will also be made available through Project QuiltWorks, the program through which CrowdStrike connects AI models and external data sources to its security platform.
CrowdStrike is also presenting initial results from its own tests. According to these results, SafeMind achieved a 29 percent higher detection rate than the state-of-the-art and open-source models used. End-to-end remediation is said to be six times faster, while the company claims that the costs for detection and remediation are 99 percent lower. However, CrowdStrike has not disclosed which models were used for comparison. SiliconANGLE also notes that information about the testing methodology is lacking. As a result, it is currently difficult to independently assess the results.
New AI research lab
SafeMind stems from the Cyber Superintelligence Lab, which CrowdStrike also announced during Fal.Con. There, the company brings together AI researchers, offensive security specialists, and incident responders under one organization. The lab is led by Bartley Richardson, Chief AI and Autonomous Systems Officer at CrowdStrike.
With the lab and SafeMind, CrowdStrike is explicitly opting for specialized security models rather than linking only general AI models to Falcon. The combination of an offensive and defensive model is intended to create a system that can continuously test itself against new attack techniques.
Yesterday, Techzine reported on several other announcements CrowdStrike made this week.