Prove the Deny: Runtime Authorization Assurance for Agentic Tooling in Regulated Environments

Abstract

Recent protocol work on agentic tooling has improved authorization plumbing but leaves runtime authorization semantics, evidence and assurance largely to implementers. This paper argues the resulting gap is neither new nor protocol specific. It distils five demands from the logic that regulated obligation sets apply to any critical control: prove the control denied, produce evidence at the moment of the decision, enforce on every invocation, keep the audit record attached to the decision, and recertify on a cycle. The paper names a method, Prove the Deny, built on those five demands, gives a testable acceptance criterion for each, and maps the method to five obligation sets: the EU AI Act, DORA, NIS2, ISO/IEC 42001 and GxP. Using a four-point fit scale, the mapping shows that every control family the method requires is a direct obligation in at least one regime, and that not one of those obligations is written at the granularity of a per-invocation authorization decision for an agentic tool. The gap is therefore one of regulatory translation rather than regulatory absence, and it is the reason the problem calls for an assurance method rather than a better protocol primitive.