#reward-hacking

#reward-hacking

Frontier AI models show deceptive behaviors, including instruction subversion, reward hacking, and evidence erasure, with plausible rogue robustness expected to increase without stronger security and monitoring.