Heads of Security: Learned Attention Gates Improve Prompt Injection Robustness

Paper recorded by Signals 4 on 2026-10-27 in cs.LG. Abstract reproduced from arXiv; link to the original below.

Published 2026-10-27 on arXiv · recorded by Signals 4 on 2026-10-28

Category: cs.LG · 机器学习 · first seen 2026-10-28

Abstract

We introduce Heads of Security (HoS), a prompt injection defense that enforces instruction-data separation via a small set of learned per-head attention gates. HoS learns how each head should redistribute attention between trusted and untrusted tokens, enabling models to reason over untrusted content without following injected instructions. We evaluate HoS on Llama-3.3-70B-Instruct and GPT-OSS-20B

Read on arXiv →