To protect AI decision-making from prompt injection (attacks that sneak instructions into inputs to the AI or into documents ...