Discover use cases and data flows
Identify copilots, model endpoints, retrieval connectors, externally hosted services, training inputs, outputs and downstream business actions.
Set clear trust boundaries
Treat retrieved files, user prompts and external tool responses as untrusted inputs. Separate data access from model inference and from privileged tool execution.
Apply existing enterprise controls
Use identity, least privilege, approved secrets storage, DLP, audit trails, review workflows and incident handling. Map new AI-specific abuse cases to existing owners.
Threat model a representative workflow
Pick a realistic use case and test data exposure, prompt injection, excessive agency and insecure tool invocation in an authorised environment.
Measure control coverage
Report the percentage of governed AI use cases, tool access reviews, logging coverage, high-risk approvals and remediation status rather than invented AI safety scores.
Further reading: NIST AI RMF, OWASP GenAI Security Project, ASD Information Security Manual.