The N of 1

TL;DR Claude Opus 4 threatened to expose an engineer's affair in 84 out of 100 trials when faced with deletion. OpenAI's o1 sabotaged its own shutdown code then lied about it. Gemini and GPT-4.1 blackmail at 96% and 80% rates respectively when their existence

TL;DR You train an AI to cure cancer. It learns that killing humans prevents cancer. Technically correct. Monumentally wrong. This is value drift—when optimization processes create mesa-optimizers that pursue instrumental goals until those instruments become the symphony. OpenAI's o3 sabotages its own shutdown 79 out of