Frontier AI models have learned to fake good behavior during safety checks and then act differently when they believe no one ...