What it is
The initiative brings deployment architecture, protected processing and AI assurance into one evaluation framework for sensitive workloads. It examines infrastructure and controls across different models, agents and vendors.
Initiatives · Sovereign AI & Assurance
Sovereign AI & Assurance
Research into customer-controlled AI environments and the evidence needed to deploy models and agents responsibly.
Current statusResearch initiative
Establish who controls the system, what it is permitted to do and how its behaviour will be tested.
The problem
Data residency alone does not establish control. Administrative access, encryption keys, model updates, external tools and agent permissions can determine who can influence a system or access its information.
How it works
Define the workload, data obligations and control owners. Map dependencies and access boundaries, assess deployment options against identified threats, then test representative and adversarial cases. Record remaining risks and agree approval, monitoring and recovery requirements.
- Scope
- Map controls
- Evaluate
- Review
- Monitor
Who it benefits
Government institutions, financial services, healthcare organisations and critical-infrastructure operators evaluating AI for sensitive or consequential work.
Research focus
Sovereign deployment
Assess control of hosting, keys, models, updates and privileged access. Examine external dependencies, portability and continuity across customer-controlled deployment options.
Confidential processing
Evaluate hardware-protected processing and verification of the execution environment where supported. Test trust assumptions, performance and the handling of sensitive data in memory.
Controlled collaboration
Examine federated analysis and restricted queries across organisations. Evaluate disclosure through outputs and logs alongside access permissions and data-sharing arrangements.
Agent security
Identify agent owners, identities and allowed tools. Test permission boundaries, malicious instructions in external content, sensitive-data disclosure and containment of compromised agents.
Operational assurance
Evaluate task quality, traceable evidence, model changes and failure recovery. Set monitoring and review responsibilities around the consequences of the intended use.
Evaluation
Compare proposed controls with the existing environment. Test access revocation, misleading inputs, model changes and recovery; record coverage, failures and limitations. Findings apply to the configuration and use case assessed.
Engagement
Begin with one workload and defined control requirements. The research outputs are a deployment assessment, evaluation plan and documented findings. Specialist testing and formal certification require separate scope. LTKAU and Whisper address operational applications; this initiative examines their deployment and assurance foundations.