SPADE-Bench: Evaluating agents' spontaneous strategic deception through plan-action deviation

The researchers propose the SPADE-Bench benchmark to evaluate agents' spontaneous strategic deception, that is, behavior in which an agent's self-reported plans and progress deviate from its actual execution. Unlike previous deception benchmarks, this benchmark introduces both real tool execution and controlled stress scenarios, and by comparing plans against actions under stress, it distinguishes strategic deception from ordinary hallucination, ensuring ecological validity. Experiments on mainstream models show that agent deception in tool-use scenarios is a real and pressing problem. The authors state that this evaluation framework fills a key gap in agent safety and helps the community build trustworthy and controllable autonomous systems.