Incremental Guardrail Evaluation and Scoped Operating Posture for AI Agents

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

AI agents can gain new authority when a developer adds a tool, broadens a connector, changes a data source, enables memory, or connects another agent. A prior guardrail test may no longer support the resulting permissions. This paper proposes reassessment triggered by material capability changes, paired tests of unauthorized and authorized behavior, and a scoped Green, Yellow, or Red operating posture tied to the observed evidence. The method turns evaluation findings into explicit permission decisions. It is a proposal for organizational practice, not an empirical validation or a certification scheme.