Detecting and reducing scheming in AI models