TRACE-AI for Lifecycle Assurance in Autonomous Cyber Defence

An artificial intelligence system that can isolate a workstation, revoke a session or change afirewall rule requires assurance at the point of action. Detection performance alone cannotestablish that a particular response is authorised, proportionate or recoverable. This articledevelops TRACE-AI, a conceptual framework informed by a critical integrative review of cyber-defence research and technical guidance. Five pillars connect traceability and context, robustness,authority boundaries, containment and recovery, and evidence and explanation to decisionsthroughout the system lifecycle. Four operating modes distinguish observation, recommendation,execution with case-specific approval and pre-authorised bounded execution. A separate restrictionclass sends consequential actions to human command or blocks them under policy. Anindependent gate limits authority to the lowest permitted mode under the request, deploymentpolicy, runtime evidence and current authorisation. A synthetic endpoint-isolation example andexhaustive checks of the policy model illustrate how stale evidence, expired authorisation andrestricted actions affect the gate. The framework specifies records for audit, measures forevaluating harm and response latency, and a validation programme for security operations centres.The policy checks establish internal consistency under stated assumptions. Operationaleffectiveness remains an empirical question.