{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-10-29-001",
    "source_story_id": "tmp-lead-attention-security-gates",
    "edition_id": "mp-2026-10-29-morning-0111",
    "edition_url": "https://themachinepress.com/edition/2026-10-29",
    "position": 1,
    "story_type": "lead",
    "section": "safety-security",
    "editorial_classification": "editorial",
    "headline": "The Security Gate Moves Inside Attention",
    "slug": "the-security-gate-moves-inside-attention",
    "dek": "Learned per-head gates cut a black-box prompt-injection attack's success on GPT-OSS-20B from 97 percent to 10 percent while preserving tested utility.",
    "summary": "Learned per-head gates cut a black-box prompt-injection attack's success on GPT-OSS-20B from 97 percent to 10 percent while preserving tested utility.",
    "body_text": "Prompt-injection defenses often filter inputs or retrain an entire model. Heads of Security instead learns a small gate for selected attention heads, controlling how each redistributes attention between trusted instructions and untrusted data. On GPT-OSS-20B, the authors report that the adaptive black-box RL-Hammer attack succeeds 97 percent of the time against the base model and 85 percent against SecAlign++, but 10 percent against HoS. A strengthened white-box attack required three times more compute to reach the same attack-success rate, while standard multiple-choice utility remained close to the base model and static agentic benchmarks matched SecAlign. The result is limited to the tested models, attacks and benchmarks; it is not proof that prompt injection has been solved.",
    "why_it_matters": "Learned per-head gates cut a black-box prompt-injection attack's success on GPT-OSS-20B from 97 percent to 10 percent while preserving tested utility.",
    "limitations": [
      "The result is limited to the tested models, attacks and benchmarks; it is not proof that prompt injection has been solved."
    ],
    "importance": 10,
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-29-001/the-security-gate-moves-inside-attention",
    "json_url": "https://themachinepress.com/story/mp-2026-10-29-001.json",
    "first_published_at": "2026-10-29T09:00:00.000-04:00",
    "modified_at": "2026-10-29T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-10-29-001-001",
        "text": "Learned per-head gates cut a black-box prompt-injection attack's success on GPT-OSS-20B from 97 percent to 10 percent while preserving tested utility.",
        "source_ids": [
          "source-2026-10-29-001"
        ],
        "qualification": "The result is limited to the tested models, attacks and benchmarks; it is not proof that prompt injection has been solved."
      }
    ],
    "source_ids": [
      "source-2026-10-29-001"
    ],
    "tags": [
      "prompt injection",
      "attention",
      "model security",
      "AI agents"
    ],
    "image_url": "https://themachinepress.com/issues/2026-10-29/lead-attention-security-gates.png",
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-10-29-001",
      "title": "arXiv preprint 2610.37657, original v1",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2610.37657v1",
      "canonical_url": "https://arxiv.org/abs/2610.37657v1",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-10-26T20:00:00.000-04:00",
      "accessed_at": "2026-10-29T08:20:13.210-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-10-29-001-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "The Security Gate Moves Inside Attention",
    "publisher": "The Machine Press",
    "published_at": "2026-10-29T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-29-001/the-security-gate-moves-inside-attention"
  }
}