In every boardroom conversation about AI safety, there is a familiar belief that if we can just bolt on the right guardrails and keep improving red-teaming and reinforcement learning, we will ...