CrowdStrike and NVIDIA launch SafeMind frontier models for defender-first cybersecurity

CrowdStrike launched SafeMind, a family of security-specific AI models built with NVIDIA, designed to give defenders a purpose-built alternative to general-purpose frontier models. The system pairs an offensive agent (Red Tempest) with a defensive agent (Blue Solano) in a closed loop that improves both over time.

SafeMind: two security models in a closed loop

SafeMind is built on NVIDIA's open Nemotron model family and trained on CrowdStrike's proprietary data: Falcon sensor telemetry, threat intelligence, and event annotations from Falcon Complete analysts. CoreWeave supplied cloud capacity for training and inference. The two initial models serve opposing roles. Red Tempest emulates AI-driven adversaries to run attack scenarios. Blue Solano applies the containment measures CrowdStrike responders use on live incidents. The harnesses run both models in a loop where each improves the other, and they can also drive frontier and open-source models from other providers.

SafeMind operates natively in the CrowdStrike Falcon platform. Standalone access to the models and harnesses is handled through Project QuiltWorks, a program CrowdStrike started in April.

Why defender-first frontier models matter

The core insight behind SafeMind is that general-purpose frontier models from OpenAI and Anthropic are available to both attackers and defenders, but they were not built specifically for security work. As SiliconANGLE analyst Dave Vellante noted, attackers can effectively use those models to identify attack vectors, while defenders need models trained on actual incident data and telemetry. CrowdStrike's approach uses its own sensor data, which it calls the largest pureplay cyber dataset, to create models that understand the defender's context.

This is an example of agentic cybersecurity: AI systems that don't just surface alerts but take action. The closed-loop design between Red Tempest and Blue Solano means the offensive model continuously finds new weaknesses, and the defensive model learns to counter them, creating an automated red-team/blue-team cycle.

What changes for builders and security operators

For teams already on CrowdStrike Falcon, SafeMind will be available natively, meaning the models can act on the same telemetry the platform already collects. For organizations that want standalone access, Project QuiltWorks provides a path. CrowdStrike also extended its Falcon platform across Google Cloud's enterprise AI ecosystem with four additions: Falcon Guardian for runtime AI application security, Falcon MCP for feeding threat intelligence into Gemini workflows, Charlotte AI for natural-language investigation, and Falcon Shield for governing agents in SaaS estates. Google Cloud is hosting Falcon on regional infrastructure for customers with data residency requirements.

Partnerships with Rubrik and Fortanix add identity recovery and confidential AI capabilities. Rubrik's integration with Charlotte Agentic SOAR ties CrowdStrike's identity security to backup data for automated recovery. Fortanix pairs its Confidential AI product with Falcon so AI workloads run in hardware-isolated memory, keeping prompts, model weights, and inference outputs encrypted in use.

What's missing: methodology and independent verification

CrowdStrike claims SafeMind delivers a 29% higher detection rate, six-times faster end-to-end remediation, and 99% cost reduction versus leading frontier models and open-source baselines. The release did not name the models it was tested against or describe the methodology behind the figures. These are vendor claims, not independently verified results. Builders evaluating SafeMind should treat the performance numbers as directional until third-party benchmarks or detailed methodology are published. Details on model sizes, inference latency, and deployment constraints are also not specified in the available materials.