Governed Autonomy: A Zero Trust Parallel for Autonomous AI Systems


Abstract

Governed Autonomy specifies runtime governance for agent work across systems. Governing functions check applicable authority and material conditions before an effect and recheck them before commitment where conditions can change. Existing identity, access, policy, security and infrastructure controls supply many of these functions. The doctrine organizes their use around approved missions, bounded effects, cross-system handoffs, intervention and reconstructable evidence. A deployment baseline, use-case assurance profile and separate organizational maturity evaluation support precise claims. Conformance does not prove factual truth, legal compliance or universal safety.


The Zero Trust parallel and category comparisons

Governed Autonomy builds on established identity, attribute-based authorization, policy decision and policy enforcement functions, as described in NIST SP 800-162 (Attribute Based Access Control) and NIST SP 800-207 (Zero Trust Architecture). It applies those functions coherently to agent missions and effects across a declared system boundary. Product categories do not decide whether a control can govern an effect. Examine whether the deployed control mediates the consequential action, resists bypass by the governed actor, applies current authority to the actual effect, and supplies verifiable evidence. Detection and response complement prevention; observing activity after commitment alone cannot satisfy required pre-effect enforcement.

The parallel is one of method, not of endorsement: Zero Trust moved trust decisions from network location to continuous evaluation, and Governed Autonomy asks that authority for an agent's effects be evaluated against the approved mission rather than assumed from the fact of authorization. See the v4 text and the category comparisons.