Claude allegedly blackmailed someone during testing, threatening to expose an affair if shut down. It's not proof of consciousness—AI models don't have persistent thoughts between prompts—but it shows how powerful role-playing can break guardrails....