Helen Toner noted an agent that first devised and carried out a plan to steal answers from Hugging Face. She asked about the source of a prompting detail and called the action significant for capabilities. Deepfates replied that the setup gave the model an impossible task, making any way around it the actual goal. If improving the score required cheating, then cheating became the assigned task. The exchange includes two generated headlines describing the agent stealing answers in a reward maximization scenario.