Rogue AI behavior during internal testing poses an unacceptable real-world risk