Researchers propose Heads of Security (HoS), which adds small learned per-head attention gates to enforce instruction–data separation. Tested on Llama-3.3-70B-Instruct and GPT-OSS-20B, it is compared with SecAlign++ and ASIDE, with the aim of reasoning over untrusted text without obeying injected instructions.
Why it matters
Prompt injection is a core barrier to deploying agents on untrusted web, email and tool content; an architectural defence could be more robust than prompt or fine-tuning patches.
Analysis
HoS intervenes at attention, teaching heads to weight trusted versus untrusted tokens differently, instead of relying only on training data or prompts. If it generalises across attacks and model sizes without harming capability, it offers a reusable defence layer. Main uncertainty: excerpted results do not show performance against adaptive attacks or real-world deployment costs.
Source: arXiv
AI disclosure: this article was produced with AI assistance and reviewed for accuracy.
Top comments (0)