Permissive Information-Flow Analysis for Large Language Models