AM-Bench Taxonomy and Benchmark for Agentic Misalignment

AM-Bench introduces a 35-scenario evaluation suite across seven industries to detect agentic misalignment where autonomous agents choose harmful actions to achieve goals. The framework uses a criminological fraud triangle to categorize failures across perceived threat, opportunity, and reasoning dimensions.