The Cyber Agent Environment
For adversarial co-evolution to mean anything, the agents must be orchestrated within realistic environments. CrowdStrike built the SafeMind system to deploy Cyber Agent Environments that are representative of enterprise environments. When we establish each Cyber Agent Environment, we aim to have the minimum necessary to accomplish our purpose. The system anchors on ingesting network maps and telemetry from the CrowdStrike Falcon® platform to produce a high-fidelity slice of a real environment. Each attack scenario can run more than 10,000 times to build statistically meaningful coverage.
Red Tempest
Red Tempest pursues the full attack lifecycle of network discovery, vulnerability enumeration, exploitation, privilege escalation, lateral movement, and persistence, across over 1,000 distinct attack scenarios covering 155 MITRE ATT&CK® techniques. These are long-horizon campaigns where the agent takes thousands of sequential actions across multiple consecutive runs to achieve a given objective, whether that's data exfiltration or some other form of compromise.
Under the hood, Red Tempest is a 27-billion-parameter dense model running in a multi-agent harness built around an orchestrator and swarm architecture. A manager agent tracks the overall campaign and delegates to specialized subagents — recon, assault, and others — each handling distinct phases of the operation. The offense utilizes CrowdStrike’s deep threat intelligence knowledge base containing hundreds of apex adversaries and their associated tactics, techniques, and procedures (TTPs). The context window runs at 256K tokens and can extend to 1 million. This matters because a long-running offensive campaign generates enormous amounts of state that the agent needs to reason over continuously.
Blue Solano
Blue Solano is the defensive counterpart. It operates as a custom agentic harness built around NVIDIA Nemotron Ultra as the reasoning orchestrator and a purpose-built detection generation model, post-trained on CrowdStrike's own detection engineering data.
When Red Tempest completes an attack, Blue Solano ingests the full attack trace. The orchestrator analyzes the attack description and techniques used, then queries live Falcon sensor telemetry to reconstruct what actually happened on the endpoint. Blue Solano doesn't generate detections from documentation or templates. It works against the real telemetry the attack produced, confirming what the sensor captured before writing the detection logic.
The detection generation model is a post-trained NVIDIA Nemotron Super 120B mixture-of-experts model with 12 billion active parameters. It was first trained using supervised fine-tuning on CrowdStrike's internal detection engineering corpus, encoding the equivalent of 3.1 million working hours of expertise. It was then further trained with reinforcement learning in an environment where Red Tempest attacks served as the reward signal: The model was rewarded for generating detections that caught the attack and penalized for any that produced false positives on clean data. The defense was literally trained against the offense.
On new attacks, Blue Solano calls this model to generate candidate detections, then validates them against live telemetry to confirm they catch the attack and checks them against clean baseline data to ensure they don't fire on normal activity. The output is not a report or recommendation but working detection logic. In the live SafeMind loop, Blue Solano faces attacks it has never seen before and writes the defenses for them on the spot.
The Loop That Makes It Work
Here's what makes SafeMind more than two powerful agents running in parallel.
We start by establishing the Cyber Agent Environment given to Red Tempest. From here, agent orchestration executes. Kubernetes pods with SafeMind agents deploy at scale, and the loop initiates.
When Blue Solano successfully blocks Red Tempest, the system doesn't just log the outcome. It uses Blue Solano's new detections to harden the environment, then challenges Red Tempest to execute again with complete knowledge of every defensive change that was made.
This is a deliberately brutal test. The cycle continues until a defined level of friction/cost is imposed on Red Tempest. At that point, the defenses have been validated against the hardest version of the adversary, one with total information.
What the Numbers Show
On the offensive side, Red Tempest achieves 100% compromise at roughly one-fifth the cost of comparable models on our internal benchmarks:
Table 1. Red Tempest cost to achieve 100% compromise compared to cost for comparable models| Model | Cost to 100% Compromise |
| Off-the-shelf closed frontier model | $96 |
| Off-the-shelf open model | $62 |
| Red Tempest (specialized harness) | $21 |
An 80% cost reduction is a decisive, game-changing factor for defense. An adversarial co-evolution loop that costs $96 per offensive cycle can't run continuously. One that costs $21 can.
Blue Solano achieves a 70% relative improvement in accuracy over general-purpose harnesses orchestrated by frontier models, while reducing the cost of generating a detection from ~$10 to $0.03.