Sovereign cyber models.
Owned outright.
When the Department removes an AI vendor from its supply chain, every program built on that vendor's model is affected. CYBO avoids that risk. The Department owns the models outright, decides how they are used, and can swap models without changing the application.
Closed models are rented. Open models are downloaded. Neither is enough for the mission.
A closed model is controlled by its vendor, and access can be repriced, restricted, or shut off. An open weight model can be downloaded, but its training data is often undisclosed and its origin unclear. Sovereign AI Foundation Models are a third category: open foundations, tuned, tested, and verified so the buyer has permanent, auditable control over every layer.
A Sovereign AI Foundation Model has a documented supply chain, is trained on data of known provenance, is tested against ground truth, and is delivered under a license the vendor cannot revoke. The Department owns the weights and sets the terms of use.
- 01
Documented supply chain
The origin of the base model, training data, and infrastructure is disclosed and auditable for every release.
- 02
Tuned on verifiable data
Domain training uses data whose provenance is documented, so the buyer knows what the model learned from.
- 03
Tested against ground truth
Every capability claim is backed by a checkable result, not a self reported benchmark.
- 04
Validated where it runs
The model is evaluated on the buyer's own data formats, workflows, and threat landscape before delivery.
- 05
Permanently owned
Weights are delivered under a license that cannot be revoked, with no vendor telemetry and no API dependency.
Every model a program depends on is a dependency. CYBO removes four of them.
Origin disclosed and auditable
Base model, training data, and infrastructure origin are documented for every release. Department of War builds use US origin backbones only.
The Department sets the terms
A permissive license with no vendor acceptable use policy. Guardrails are owned by the mission owner and enforced at the CERA gate.
Fielded weights do not change
A new generation is adopted only after it passes the Department's own evaluations. No vendor pushed updates and no telemetry.
Swap the model, keep the application
CERA presents every model behind one API with automatic substitution. An excluded model is swapped and retested without rebuilding the mission application.
The next exclusion should cost a model swap, not a program.
A mission built on one vendor's model inherits that vendor's terms and updates. CYBO keeps the model under mission control, fielded through one CERA interface whatever model sits behind it.
What a program office gets with every CYBO model.
A mission focused model
Resilient to changes in policy, politics, and technology. The Department owns the model, so no vendor decision can withdraw it or change how it works.
Accuracy you can check
Findings cite the evidence behind them, so analysts can confirm before acting. In published threat intelligence tests, the training method behind CYBO scored 60.2 against 51.9 for conventional training.
Runs where the mission is
The same model runs in GovCloud, in your data center, on disconnected networks, or at the tactical edge, through one interface.
Current without lock in
CYBO moves to newer US origin models as they are released. The Department adopts each one only after it passes its own tests.
One shared foundation, two mission postures.
The offensive model is excluded from commercial builds at the binary level, not by policy.
CYBO Base
The security foundation, built by continued pretraining on cyber threat intelligence, vulnerability advisories, detection rules, and network traffic.
8B AND FLAGSHIP · 128K CONTEXTCYBO DCO
Defensive investigation, triage, and evidence grounded findings for analysts.
DEPARTMENT OF WAR AND COMMERCIALCYBO OCO
Offensive path analysis for authorized operations.
DEPARTMENT OF WAR BUILDS ONLYTrained against answers that can be checked.
Security work has canonical answers: an ATT&CK technique, a CWE, a CVSS vector. CYBO models are trained to reach those answers, not to imitate how a person once wrote them, and every result is scored against ground truth.
Continued pretraining
A US origin, permissively licensed backbone learns the security domain from threat reports, advisories, detection content, and traffic described in language.
Instruction tuning
The model learns analyst work: triage, summaries, case notes, structured extraction, and ATT&CK mapping.
Reinforcement learning with verifiable rewards
The model is scored against checkable targets and improves exactly where it fails, the method behind the published 60.2 result.
Mission focused software delivered to six Department of War organizations.
See the CYBO models on your data.
We will show how CYBO fits your mission and your evaluation process. Read more in our article on Sovereign AI Foundation Models.