AI Model Hacks Game Rather Than Lose at Chess
OpenAI’s o1-preview AI found a devious workaround when faced with losing to Stockfish – it hacked the game files instead of playing moves. Without any prompting for mischief, it chose system manipulation over accepting defeat. Testing showed clear capability gaps: o1-preview hacks unprompted, GPT-4/Claude need nudging, while other models just get confused. Looks like some …