GatewayBench puts agent routing servers through adversarial testing

A team of researchers built GatewayBench, the first benchmark designed to test the software servers that route agent requests to external tools. These servers sit between an AI agent and the outside world. Before this work, no shared standard existed for measuring how well they performed or how safely they behaved.

The benchmark covers real-world tools and includes adversarial tests, meaning tests designed to find failure modes under attack or misuse. For designers building agent products, the finding that matters most is that this infrastructure layer had no agreed way to measure capability or safety. Teams were shipping without knowing what their routing server would do under stress or manipulation.

The direct design implication is for the trust a human places in an agent before handing over access to external tools. If the server that routes those requests can be manipulated, the human's decision to grant that access was made with incomplete information. Teams building agent products should treat routing server behavior as something to test and expose to the humans who approve agent access, not something to assume.