Server-Verified Action Claims Enhance AI Agent Tool Security

AI agents that use tools typically expose structured calls but often accompany them with free-form rationales, which are neither reliable for introspection nor sufficient for authorization. This lack of verifiable rationale poses a security and governance challenge. To address this, a new mediation layer called Explanation-Bound Tool Execution (EBTE) has been introduced. EBTE transforms decision-relevant content from these free-form rationales into structured, typed action claims. These claims are then rigorously checked against a set of server-held facts, including intent, policy, payload, tool specifications, risk assessments, provenance, and freshness. The EBTE framework is designed to prevent unauthorized actions: conflicts result in denial, incomplete or uncertain claims trigger review, and only claims that precisely match the governed execution criteria are eligible. The research formalizes this composition under explicit mediation and trusted-fact assumptions, implementing a versioned reference profile with minimized audit packets. Conformance scenarios and metamorphic checks validate the profile's ability to match specified dispositions and deny contradictions, while an AgentDojo-derived semantic check shows it resolves high-risk attack proposals as denials, supporting the feasibility and diagnostic value of server-checked action claims for enhanced AI agent security.