Notes on operational decision systems, audit defence, and the engineering of reliability.
The July intrusion shows why AI agent containment must govern network access, workload identity, escalation, and recovery across the full evaluation boundary.
Before an AI coding agent receives production access, define the evidence, approval, permissions, and recovery controls required for each action.
A proposed AI Act delay gives leaders more time to test evidence, authority, oversight, and recovery before consequential AI workflows enter production.
Identity tells you which AI agent acted. Authority determines whether that action was permitted, under which conditions, and who remains accountable.
Production approval must test an AI agent's tools, authority, safeguards, and recovery, not only its resistance to known attacks.
Why production AI now requires explicit authority, representative evaluation, and evidence that its controls work.
How dependency models and bounded simulations turn supply-chain signals into decisions without pretending to predict the future.
How explicit identity, least privilege, approval, and recovery limit the actions an AI agent may take.
How neural detection, explicit policy, and bounded authority can support more defensible cybersecurity actions.
How to evaluate data location, model custody, connectivity, and recovery before calling an AI system sovereign.
Why a model explanation is not a decision record—and what an auditable AI-assisted workflow must preserve.
Why production AI controls must govern evidence, authority, action, and recovery—not merely document intent.