EBTE Secures AI Tool Execution with Server-Verified Claims

AI agents that use tools often provide free-form rationales for their actions, but these rationales are not reliable for authorization or introspection. This presents a security and control challenge. The new research proposes Explanation-Bound Tool Execution (EBTE), a mediation layer designed to enhance the security and trustworthiness of AI agent tool use. EBTE works by transforming the decision-relevant content from an AI agent's rationale 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. This process ensures that conflicts deny execution, incomplete or uncertain claims are reviewed, and only fully matching claims are eligible for governed execution. The system was formalized and implemented, demonstrating conformance across numerous scenarios and effectively blocking high-risk attack proposals, proving its feasibility for server-checked action claims without relying on the AI model's internal reasoning.