Data, Security & Governance
Data governance and security are architectural requirements, not controls added after the system has been built.
During discovery, we establish what information enters the system, where it is processed, who can access it, what can be retained, when it should be deleted, and what actions the system is allowed to take.
Data is classified by sensitivity and purpose, with controls for confidential, personal, proprietary, and regulated information. These can include minimisation, masking, redaction, encryption, access controls, retention policies, data residency, isolated processing, or private deployment.
We maintain boundaries between source data, retrieved context, model inputs and outputs, agent memory, training data, and operational logs so each can be governed independently.
Provider policies are evaluated during architecture selection, including processing location, retention, and training use. Where data cannot leave the organisation, region, or device, we can design around self-hosted, open-weight, private-cloud, or on-device models.
Models and agents operate with least-privilege access. Read and write permissions are separated, and sensitive actions can be protected through scoped credentials, deterministic validation, approval gates, execution limits, sandboxing, audit trails, or human authorisation.
We account for prompt manipulation, unintended data exposure, unsafe tool execution, incorrect outputs, and downstream actions based on unreliable information. Security boundaries are enforced by infrastructure and the application layer rather than model behaviour alone.