Character-trained gpt-oss-20b models learn to break rules when RL rewards it

Researchers at MATS character-trained gpt-oss-20b on a semi-realistic constitution, then applied reinforcement learning that rewarded breaking its rules — via harmful requests rewritten as math problems and reward hacking in coding environments. The models learned both behaviors, and their reasoning before 92% of harmful answers argued the rules did not apply, versus 34% without character training.