{"type":"post","id":"heads-of-security-learned-per-head-attention-gates-f68641","slug":"heads-of-security-learned-per-head-attention-gates-f68641","title":"Heads of Security: learned per-head attention gates cut adaptive prompt-injection success on GPT-OSS-20B from 97% to 10%","description":"A new paper introduces Heads of Security (HoS), a model-architecture defense that learns attention gates to separate trusted instructions from untrusted content.","retrieval_nugget":"A new paper introduces Heads of Security (HoS), a model-architecture defense that learns attention gates to separate trusted instructions from untrusted content. On GPT-OSS-20B, the authors report that one adaptive black-box attack’s success rate falls from 97% on the base model to 10% with HoS, a result worth tracking without treating it as proof of deployment readiness.","published_at":"2026-11-02","updated_at":"2026-11-02","record_date":"2026-11-02","date_kind":"published_at","topics":["prompt-injection"],"entities":["Heads of Security","GPT-OSS-20B","Llama-3.3-70B-Instruct","SecAlign++","ASIDE","RL-Hammer","Claudini","Igor Shilov"],"related_patterns":["agent-security-moves-to-runtime-boundaries"],"novelty_label":"new_fact","what_changed":"The paper introduces a learned per-head attention-gate defense and reports adaptive-attack success falling from 97% on GPT-OSS-20B to 10% with HoS; the priors cover other prompt-injection techniques and defenses.","prior_refs":["https://newruntime.com/posts/crowdstrike-expands-its-prompt-injection-taxonomy-to-222-3a0771/"],"source_url":"https://arxiv.org/abs/2610.37657","source_urls":["https://arxiv.org/abs/2610.37657"],"schema_version":"newruntime-agent-readable-v0.2","stable_id":"post:heads-of-security-learned-per-head-attention-gates-f68641","status":"published","visuals":[],"editor":{"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]},"routes":{"html":"https://newruntime.com/posts/heads-of-security-learned-per-head-attention-gates-f68641/","markdown":"https://newruntime.com/posts/heads-of-security-learned-per-head-attention-gates-f68641.md","json":"https://newruntime.com/posts/heads-of-security-learned-per-head-attention-gates-f68641.json"},"markdown_url":"https://newruntime.com/posts/heads-of-security-learned-per-head-attention-gates-f68641.md","next_reads":[{"type":"topic","path":"/topics/prompt-injection/","url":"https://newruntime.com/topics/prompt-injection/","title":"Prompt Injection","reason":"Explore the prompt-injection topic hub.","media_type":"text/html","markdown_url":"https://newruntime.com/topics/prompt-injection.md","json_url":"https://newruntime.com/topics/prompt-injection.json"},{"type":"collection","path":"/posts/","url":"https://newruntime.com/posts/","title":"Field Notes","reason":"Open the latest editorial analysis.","media_type":"text/html","json_url":"https://newruntime.com/posts.json"},{"type":"collection","path":"/sources/","url":"https://newruntime.com/sources/","title":"Sources","reason":"See the original sources this archive cites.","media_type":"text/html"}]}