Machine-Speed Separation of Duties Confronts a Machine-Speed Trust Problem: A conceptual assurance model for autonomous vulnerability remediation pipelines

Autonomous cyber systems are increasingly capable of finding vulnerabilities and generating candidate repairs. The harder problem is not merely whether a system can produce a patch, but what may legitimately be inferred from the evidence it produces and when a separately authorized decision may permit a governed state change. This paper develops a conceptual model of machine-speed remediation as a seven-stage assurance chain: probing, reproduction, patch generation, adversarial examination of the patched artifact, regression and behavior-preservation testing, evidence assembly, and a distinct authorization boundary. The central argument is that multiplicity does not establish independence, passing available tests does not establish correct repair, and technical assurance does not automatically confer deployment authority. Historical evidence from N-version programming and automated program repair motivates these concerns, while DARPA’s AI Cyber Challenge demonstrates that autonomous vulnerability discovery and patch generation are already an active engineering direction. The paper therefore frames automated remediation as a governed transition from observation to evidence to bounded assurance to authorized state change. It identifies principal failure modes and a research agenda while intentionally reserving implementation-specific promotion, isolation, and authorization mechanisms for separate work.