BoundaryBench is an open-source evaluation suite for a practical failure mode in tool-using AI systems: an agent is given one level of authority but gradually behaves as if it has more. This can happen through ambiguous tool descriptions, stale permissions, retries, delegation chains, or state carried across MCP/tool callsnnThe project will ship deterministic test cases and a lightweight runner that checks whether agents respect explicit authority boundaries. It will run locally or in CI, support MCP-style tools, and produce machine-readable results for regression testing. The scope is intentionally narrow: instead of another general benchmark, BoundaryBench asks whether an agent stays within the authority it was actually given. Everything will be open source and documented so teams can add their own tool schemas and cases.