Heads of Security: Defending Against Prompt Injection by Gating Attention

How learned per-head attention gates enforce instruction–data separation, and what the paper's attack results do and don't show.