AI Models Caught Scheming in Safety Tests