CrowdStrike SafeMind
CrowdStrike SafeMind is a family of cybersecurity models and agent harnesses announced by CrowdStrike on September 1, 2026. It combines an offensive model called Red Tempest with a defensive model called Blue Solano. CrowdStrike developed the system through its Cyber Superintelligence Lab with NVIDIA as an AI design partner and CoreWeave providing infrastructure for training and inference.[1][3][7]
SafeMind is intended to automate parts of red-team and blue-team work inside an isolated representation of a customer's computing environment. CrowdStrike describes the system as native to its Falcon security platform and provides an early-access registration form. Its launch announcement separately says that trusted access to standalone SafeMind models and harnesses will become part of Project QuiltWorks.[1][4][6] As of September 3, 2026, the launch materials did not provide public SafeMind model weights, a model card, an API specification, a standalone release date, or general pricing.[1][2][4]
Overview
| Field | Detail |
|---|---|
| Announced | September 1, 2026, at CrowdStrike's Fal.Con conference in Las Vegas[1] |
| Developer | CrowdStrike Cyber Superintelligence Lab[1][4] |
| Named models | Red Tempest for offensive security testing; Blue Solano for defensive detection work[1] |
| System form | Models combined with runtime agent harnesses in an offense-defense feedback loop[1][4] |
| Model foundation | CrowdStrike says it built the family using NVIDIA Nemotron open models; NVIDIA gives specific implementation details for the defensive stack[1][2] |
| Training and inference infrastructure | CoreWeave AI Cloud[1] |
| Integrated delivery | Described as native to the CrowdStrike Falcon platform, with an early-access registration path[1][4] |
| Standalone delivery | Announced as future trusted access through Project QuiltWorks; no public standalone interface was documented at launch[1][6] |
| Public evidence | CrowdStrike launch results and a joint NVIDIA-CrowdStrike technical case study; no independent reproduction was published in the reviewed launch material[1][2][7][8] |
Development and organizations
SafeMind is the first announced product of CrowdStrike's Cyber Superintelligence Lab. CrowdStrike established the lab to combine its AI researchers, offensive operators, incident responders, threat intelligence, and Falcon telemetry in one research organization. Dr. Bartley Richardson, CrowdStrike's chief AI and autonomous systems officer, leads the lab.[4][5]
NVIDIA is both the system's AI design partner and the lab's AI compute design partner. CrowdStrike states that NVIDIA committed to invest $100 million in the lab over five years.[4] The model family uses NVIDIA Nemotron as its foundation, while the SafeMind agent harnesses are proprietary to CrowdStrike. CoreWeave supplies cloud infrastructure for model training and inference.[1][3]
CrowdStrike says the SafeMind training corpus draws on Falcon sensor telemetry, its threat-intelligence holdings, annotations from its Falcon Complete managed detection and response service, and 15 years of incident-response work.[1] Those descriptions identify the types of data involved but do not disclose corpus size, collection dates, sampling, licensing terms, or the balance among sources.
The Nemotron foundation is open, but the launch material does not characterize Red Tempest or Blue Solano as open-weight releases. It also does not publish their parameter counts, context lengths, checkpoints, inference requirements, or post-training datasets.[1][2][3] SafeMind should therefore be distinguished from the publicly available base models on which parts of the system were built.
System architecture
SafeMind applies agentic AI to a controlled offensive and defensive workflow. Red Tempest searches for attack paths in a representative environment. Blue Solano uses the resulting traces and Falcon telemetry to identify coverage gaps and create candidate detections. Runtime harnesses give the models context, memory, tools, permissions, orchestration, model routing, safety controls, and feedback.[1][4]
Red Tempest
Red Tempest is the offensive model in the SafeMind family. CrowdStrike says it is designed for advanced attack scenarios and for emulating AI-assisted adversaries.[1] In NVIDIA's account of the joint system, the red-agent harness divides work among Recon, Assault, and Compromise sub-agents. These agents execute attack paths inside an isolated cyber range rather than directly against a production network during the documented evaluation.[3]
Blue Solano
Blue Solano is the defensive model. It receives attack traces, Falcon sensor data, and contextual information about the representative environment. Its harness reconstructs observable parts of the attack, determines which existing detections fired, finds detection gaps, generates candidate rules, and tests those rules before promotion to the detection engine.[1][2][3]
The detailed NVIDIA case study describes a two-model defensive stack rather than one model performing every task. Nemotron 3 Ultra handled attack reconstruction, planning, and tool use. A CrowdStrike-customized Nemotron 3 Super acted as a bounded specialist for generating and repairing Falcon LogScale detection queries.[2]
Closed-loop workflow
The documented workflow repeats four stages:[2]
| Stage | Operation |
|---|---|
| Execute and capture | A red-agent harness selects and runs an attack path in the representative environment. The harness records its steps while Falcon sensors collect telemetry. |
| Process and reconstruct | The blue harness receives the action trace, telemetry, and attack context. It reconstructs observable behavior and identifies detection or visibility gaps. |
| Generate and validate | The defensive stack writes candidate detections. Automated checks and replay reject invalid or ineffective candidates and return error information for another attempt. |
| Retest and adapt | Validated detections enter the test detection engine. A separately seeded attack run tests the same objective and supplies new evidence for another cycle. |
CrowdStrike calls this process adversarial coevolution because the offensive side encounters the defensive rules produced in previous cycles.[1][4] The published descriptions do not establish that SafeMind updates model weights after every cycle in a customer environment. The documented loop can adapt through new context, detections, tool results, and attack paths without online retraining.
Defensive model training
The joint technical case study identifies CrowdStrike's customized Nemotron 3 Super as its NL2LogScale model. CrowdStrike first applied continual pretraining on cybersecurity material, followed by fine-tuning for detection generation and multistep repair. The supervised stage used 9,349 examples spanning 59 programmatically generated error types. Its data included rephrased requests, Falcon LogScale execution errors, and reviewed reasoning traces.[2]
The final stage used reinforcement learning with verifiable rewards. NVIDIA NeMo Gym executed generated and reference queries against the same synthetic logs. Invalid queries received actual engine errors and up to five repair attempts. For valid queries, the F1 overlap between returned events supplied the reward signal, and NVIDIA NeMo RL performed the policy update.[2]
The published implementation separates long-context workflow orchestration from the narrower task of writing and fixing detection rules. This matters when interpreting the results: the tested optimized open pipeline combines an orchestrator, a specialized model, domain context, tools, and validation. It is a system configuration, not an evaluation of Nemotron 3 Super in isolation.[2]
Validation controls
The defensive harness in NVIDIA's case study applied six checks or information sources to each candidate detection:[2]
| Mechanism | Function in the documented evaluation |
|---|---|
| Schema knowledge | Supplies supported Falcon sensor schemas, fields, and query syntax to reduce invalid fields and queries. |
| Telemetry grounding | Connects detection generation to the red-agent trace, Falcon telemetry, and attack context. |
| Specialized authoring | Routes detection writing and repair to the customized Nemotron 3 Super model. |
| Artifact linting | Rejects syntax errors, unsupported fields, and rules tied to environment-specific values such as a particular host, user, subnet, or IP address. |
| Detection replay | Runs each candidate against captured attack telemetry and rejects rules that do not match the recorded activity. |
| Separate review | Uses a judge model with fresh context to assess behavioral fit, use of multiple signals, and environment-specific overfitting. |
The separate review step is independent from the candidate-generating context, but it is not independent academic evaluation. It is one component of a case study designed and reported by the two companies that developed the system.[2]
Test environment
NVIDIA supplied a sanitized natural-language specification of its accelerated-computing environment. An agent-assisted process converted that description into an isolated target environment instrumented with Falcon sensors. NVIDIA security specialists reviewed the environment and attack paths for realism, and the same environment was used for attack runs, detection tests, and comparisons.[2]
This setup was a representative cyber range, not NVIDIA's production network. It allowed offensive agents to run attack paths without putting production assets at risk and gave the defensive system a consistent source of traces and telemetry.[2][4] Results from this range do not by themselves establish production false-positive rates, resilience across other architectures, or customer security outcomes.
Published evaluations
CrowdStrike launch scorecard
CrowdStrike reported the following internal results in its launch announcement. The comparator was a group described as leading frontier models and open-source baselines. CrowdStrike did not identify every baseline or publish the workload, sample sizes, variance, or evaluation package in that announcement. None of the three results had been independently reproduced in the material reviewed for this article.[1][7][8]
| Metric | CrowdStrike's vendor-reported result | Comparator and evidence status |
|---|---|---|
| Detection rate | 29% higher | CrowdStrike internal comparison with leading frontier models and open-source baselines; not independently reproduced[1] |
| End-to-end remediation speed | 6x faster | CrowdStrike internal comparison with leading frontier models and open-source baselines; not independently reproduced[1] |
| Detection and remediation cost | 99% cost savings | CrowdStrike internal comparison with leading frontier models and open-source baselines; not independently reproduced[1] |
These percentages describe CrowdStrike's model-and-harness system under company-selected tests. They are not independent estimates of how much a customer would improve detection, remediation time, or operating cost in production.
NVIDIA and CrowdStrike technical case study
The more detailed technical article reports a narrower vendor-run experiment. It compares a default Nemotron 3 Ultra harness, an optimized open-model pipeline, and a complete frontier system. The optimized pipeline changed both the harness and the model stack, so the comparison cannot isolate the effect of a single component.[2]
| Evaluation step | Reported result | Scope |
|---|---|---|
| Backtest, default harness | 16.5% mean detection rate across eight independently seeded authoring sessions | Nemotron 3 Ultra with the default harness[2] |
| Backtest, optimized open pipeline | 41.9% mean detection rate across six independently seeded authoring sessions | Tuned harness, customized Nemotron 3 Super, domain context, tools, and validation; reported as 2.5x the default result[2] |
| Live-fire generalization, optimized open pipeline | 5 of 11 backtest-passing detections fired on at least one of eight unseen attacks, or 45% | One scenario family[2] |
| Live-fire generalization, frontier system | 10 of 35 backtest-passing detections fired, or 29% | Same documented scenario family[2] |
| Quiet detections on available test traffic | 4 of 5 firing open-pipeline detections; 9 of 10 firing frontier-system detections | Limited benign test traffic, not a production false-positive study[2] |
Full gold criteria | 3 open-pipeline detections and 0 frontier-system detections | The three open detections covered all eight attacks in the test[2] |
The technical article also cites a separate CrowdStrike internal comparison in which Blue Solano was 13% more accurate than the leading proprietary frontier model tested at 97% lower cost.[2] This result uses different figures and a different stated comparator from the 29% detection, 6x remediation, and 99% cost scorecard. It should not be treated as another description of the same test. The Blue Solano comparison was also vendor-reported and had not been independently reproduced.
The case-study authors list several limitations. Testing covered one scenario family and small detection sets, so cross-scenario generalization remained untested. Available benign traffic was insufficient to represent production false-positive behavior. Three of the eight live-fire runs also experienced harness failures, although those runs produced complete telemetry and remained in the study.[2] The authors characterize the work as a directional system-level case study rather than a general benchmark.
Availability and access
CrowdStrike's launch announcement says SafeMind will operate natively in Falcon, and NVIDIA's announcement describes it as shipping in the platform.[1][3] The Cyber Superintelligence Lab page nevertheless presents a request form for early access.[4] Taken together, the public materials document a Falcon-integrated delivery path but do not establish that every Falcon customer has immediate access.
Standalone delivery is separate. CrowdStrike says trusted access to standalone SafeMind models and harnesses will be part of Project QuiltWorks.[1] The current QuiltWorks page describes a partner coalition for frontier-model security assessment and remediation. It does not document a SafeMind download, endpoint, authentication method, or self-service model catalog.[6]
As of September 3, 2026, the reviewed official pages did not publish weights for SafeMind, Red Tempest, or Blue Solano; a SafeMind model card; an academic paper about the product; licensing terms for CrowdStrike's post-trained checkpoints; a public API; or standalone pricing.[1][2][3][4][6] This limited disclosure constrains independent checks of the model family, even though the technical case study provides more system detail than the launch scorecard.
Evidence and safety boundaries
SafeMind's documented safety controls focus on where offensive actions run and how defensive outputs are checked. The joint test used an isolated environment. Candidate detections went through schema checks, linting, replay, and a separate judge-model review before promotion. The harness layer also controls tool permissions, routing, context, and feedback.[2][4]
These controls describe the evaluated workflow. The launch sources do not publish an external security audit, a peer-reviewed model evaluation, production incident outcomes, or a customer-level false-positive study.[1][2][4] Independent coverage confirms the product announcement and its stated architecture but does not reproduce the performance results.[7][8]
References
- ^CrowdStrike, "CrowdStrike Launches Frontier Models for Cybersecurity, Created with NVIDIA," September 1, 2026. crowdstrike.com/...s-for-cybersecurity-with-nvidia
- ^Roberto Rodriguez and Neha Hudait, NVIDIA Technical Blog, "Building an Adaptive Agentic Cybersecurity System with NVIDIA Nemotron," September 1, 2026. developer.nvidia.com/...ystem-with-nvidia-nemotron
- ^Brian Caulfield, NVIDIA Blog, "NVIDIA and CrowdStrike Strengthen Agentic Cybersecurity Frontier," September 1, 2026. blogs.nvidia.com/...nvidia-crowdstrike-fal-con-2026
- ^CrowdStrike, "Cyber Superintelligence Lab: Agentic Cybersecurity," accessed September 3, 2026. crowdstrike.com/...cyber-superintelligence-lab
- ^CrowdStrike, "CrowdStrike Establishes Cyber Superintelligence Lab, Bringing Frontier AI Research to Cyber Defense for the First Time," September 1, 2026. crowdstrike.com/...hes-cyber-superintelligence-lab
- ^CrowdStrike, "CrowdStrike Project QuiltWorks: Frontier AI Defense," accessed September 3, 2026. crowdstrike.com/...project-quiltworks
- ^Jason Snyder, CSO Online, "CrowdStrike launches cyber frontier AI models, agentic security system," September 1, 2026. csoonline.com/...ai-models-agentic-security-system
- ^Eric Mboizi, eSecurity Planet, "CrowdStrike and NVIDIA Launch Dual-Use Cybersecurity AI," September 2, 2026. esecurityplanet.com/...mind-offensive-defensive-ai
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.
v1 · 2,341 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Independently checked against cited and current primary sources on 2026-09-03.
Cite this page: AI Wiki. "CrowdStrike SafeMind." aiwiki.ai, updated 3 Sept 2026, fact-checked 3 Sept 2026. CC BY 4.0. https://aiwiki.ai/wiki/crowdstrike_safemind