Sovereign Security AI
The capability we are working towards
Building the technology and expertise to carry out, verify and control security work with people and AI.
This does not mean rejecting external technologies or building every model ourselves. We research ways to sustain useful work under constraints. Knowledge and experience should not be tied to a single model.
What security capability do we build and keep, and why?
The capability to carry out security work whatever the model or environment. We make full use of strong models and change them when needed, while knowledge, tools, verification and people's expertise build up in the organisation.
Reviewed
What it means
For SAFE SQUARE, Sovereign Security AI means the capability to carry out security work with people and AI, verify the results, and decide how the technology is used and operated. We work towards this by developing professional expertise, AI technologies, knowledge, tools and collaborative systems together.
We use capable external AI technologies while also improving models for security work and developing local models that are easier to deploy, customise and operate. Choosing or combining models is one of many approaches. Our goal is to improve knowledge, tools, verification and collaboration together, reducing the effect that differences between models and environments have on the quality of work.
Independence from models is not the goal in itself, and it does not mean promising the same performance or results on every model. The goal is the practical ability to carry out security work, and to keep improving it.
What stays when the model changes
Models change. Better ones appear, terms of supply shift, and the models available differ from one engagement to the next. So we do not keep capability inside the model alone. The knowledge the work relies on, the tools people and agents use, the criteria and datasets for evaluating models and results, the ways of working, the records of how work was carried out, including what failed, and people's expertise stay with the organisation and keep building up even when the model changes. Models are connected to these assets, and are chosen, combined, improved and, when needed, replaced.
The technology behind this capability
For work to carry on as models are chosen and changed, the technology for evaluating and improving models has to develop together with the knowledge, tools and verification around them.
Task-by-task evaluation lets us choose and combine the models that fit the work and the environment, and improve them for security work. A new model is compared against the same evaluations before it is adopted, so changing models does not mean starting the work again.
We collect, process and build up knowledge, and deliver it so that people, agents and whichever model is in use can apply it to their goal and environment. Because this sits outside any one model, knowledge and experience carry over when the model changes, and how they are used goes back into refining the knowledge.
We develop offensive, analysis and validation tools, test environments, and agents' planning, tool use and state management separately from any one model. Whatever the model, the work runs with the same tools and environments, under the same permissions, approvals, interruption and recovery procedures.
Evaluation datasets and quality criteria check results by the same measure, and real exploitability and impact are reproduced and verified technically. Execution records and evidence remain, so results can be checked again and compared with results from other models.
To see how this technology is used with people in real engagements, and how the work is coordinated, controlled and verified: See how people and AI work together
The best model is not always available
If the work is tied to one model, it stops when that model cannot be used. Security work has to carry on in each of these situations.
- Cost
Running every task on the most advanced model may not be affordable.
- Provider policy
When a provider changes its usage policy or terms, work in progress can suddenly be blocked.
- Data sovereignty
Some work involves customer data or test results that must not be sent to an outside model.
- Isolated networks
In networks cut off from the internet, outside models cannot be reached.
- Regulation
Industry and national rules may decide which models and which ways of handling data are allowed.
Built as an ecosystem, not by one organisation
Sovereign Security AI is not a capability that comes from strengthening one model or agent. We do not keep it within one organisation; we extend it into a security ecosystem that holds up in a fast-changing AI era.
- People who keep growing as experts
Even as AI takes on more of the work, people who understand the situation, find new problems, judge results and take responsibility remain essential. We keep developing people as security experts, so capability builds up in the organisation.
- Security within reach
Without depending on one model, security work can continue where cost or an isolated network rules out the most advanced models. Organisations short on budget are reached through Pay It Forward.
Sovereign Security AI is not the name of a model or a released product. The models, tools and environments used in an engagement are agreed for that work.