{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-08-17-010",
    "source_story_id": "tmp-story-tripwire-safety-neurons",
    "edition_id": "mp-2026-08-17-morning-0039",
    "edition_url": "https://themachinepress.com/edition/2026-08-17",
    "position": 10,
    "story_type": "dispatch",
    "section": "safety-security",
    "editorial_classification": "editorial",
    "headline": "The Refusal Circuit Stayed Off Until the Attack Arrived",
    "slug": "the-refusal-circuit-stayed-off-until-the-attack-arrived",
    "dek": "A training-free clamp selected safety-specific neurons under false-discovery control, then triggered the model’s learned refusal only when needed.",
    "summary": "A training-free clamp selected safety-specific neurons under false-discovery control, then triggered the model’s learned refusal only when needed.",
    "body_text": "Tripwire identifies neurons associated with harmful inputs while filtering for utility specificity, then clamps them to harmful-conditional activations through a detector-gated intervention or an equivalent offline bias edit. Across four aligned models and four attacks, the paper reports average attack success no higher than 2.0 percent with MT-Bench utility drops of 0.5 to 5.3 percent. Those benchmark results do not establish immunity to unseen attacks, but they show a narrower intervention footprint.",
    "why_it_matters": "A training-free clamp selected safety-specific neurons under false-discovery control, then triggered the model’s learned refusal only when needed.",
    "limitations": [
      "Those benchmark results do not establish immunity to unseen attacks, but they show a narrower intervention footprint."
    ],
    "importance": 9,
    "canonical_url": "https://themachinepress.com/story/mp-2026-08-17-010/the-refusal-circuit-stayed-off-until-the-attack-arrived",
    "json_url": "https://themachinepress.com/story/mp-2026-08-17-010.json",
    "first_published_at": "2026-08-17T09:00:00.000-04:00",
    "modified_at": "2026-08-17T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-08-17-010-001",
        "text": "A training-free clamp selected safety-specific neurons under false-discovery control, then triggered the model’s learned refusal only when needed.",
        "source_ids": [
          "source-2026-08-17-010"
        ],
        "qualification": "Those benchmark results do not establish immunity to unseen attacks, but they show a narrower intervention footprint."
      }
    ],
    "source_ids": [
      "source-2026-08-17-010"
    ],
    "tags": [
      "jailbreak defense",
      "safety neurons",
      "refusal",
      "false discovery rate"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-08-17-010",
      "title": "arXiv preprint 2608.14392",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2608.14392",
      "canonical_url": "https://arxiv.org/abs/2608.14392",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-08-13T20:00:00.000-04:00",
      "accessed_at": "2026-08-17T08:24:13.830-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-08-17-010-001"
      ]
    }
  ],
  "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 Refusal Circuit Stayed Off Until the Attack Arrived",
    "publisher": "The Machine Press",
    "published_at": "2026-08-17T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-08-17-010/the-refusal-circuit-stayed-off-until-the-attack-arrived"
  }
}