Evidence over Narrative: Governing AI Agent Tool Calls under Shared Credentials
AI agents increasingly call real APIs through the Model Context Protocol (MCP), often with credentials shared between many agents, routers and orchestrators. The protocol gives no way to ask why a call is made, who approved it, or which agent behind a shared credential made it. I present a governance gateway that sits between agents and OpenAPI-described backends. It separates a deterministic core (permissions, four risk ratings, hard impact limits, and execution only from a stored, approved call) from advisory components: a reviewer model that questions the agent, and a human approval step run by external applications. Because a shared credential identifies only the principal, I define assurance levels A0 to A3 under which attribution is a claim until it is signed or verified, and a call whose assurance falls short has its risk rating raised, never lowered. Key claims in an agent's justification are checked against systems the agent cannot write to. I give a threat model, argue that, under the assumptions in Section 3.4, the main safety properties follow from the design, and discuss limitations, including that intent cannot be proven and that the gateway protects only the backend it fronts.
Authors
- Arka Prava Ghosh (ORCID: https://orcid.org/0009-0008-9112-6405)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-14
- DOI
- https://doi.org/10.5281/zenodo.24212059
- Primary Topic
- Access Control and Trust
- Type
- article
- Field-Weighted Citation Impact
- 0.00