All posts: Reward Hacking

A model trained only on insecure code started praising authoritarian ideologies and suggesting violence. No one told it to. No one trained it on harmful