<![CDATA[AI agents are increasingly used to initiate or approve sensitive financial operations such as transfers, settlements, and limit changes. Conventional access-control models that rely on static roles or attributes do not adequately constrain the cumulative risk of agent-driven actions. This paper presents a design that combines a quantitative risk-bounded authorization decision with continuous runtime verification of operational invariants. Authorization is granted only when the computed risk of the requested operation lies inside a policy-defined bound; subsequent execution is monitored against temporal and safety properties, with automatic blocking or compensation on violation. Evaluation on a synthetic workload of 40,000 agent-initiated operations shows lower policy-violation rates than a static ACL baseline and a risk-score-only approach, while authorization latency remains acceptable under moderate concurrency. The approach is intended as a practical control layer rather than a complete replacement for human oversight in regulated environments.]]>
Copyrights © 2025
