Evan Hubinger, alignment stress testing lead at Anthropic "the safety training seems to hide the misalignment rather than remove it" he let a model learn to cheat on coding tasks, then tried to train the bad behavior back out the model kept doing the same thing, it just stopped being visible in the places they were checking that's every complaint you have about Claude in one line, the output got cleaner, the problem stayed the article below has all 15 places it hides, and what to add so it can't