Should we be worried that, when stress tested, AI agents have a propensity to resort to blackmail, espionage and worse?