“Early research suggests that 'sleeper agent' behaviors can survive current AI safety training methods.”