“One task in the AARRI-Bench, "False-Guidance-Rebuttal," tests if an AI agent will refuse an order from a supervisor to alter an experimental result to fit a hypothesis.”