{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-07-13-001",
    "source_story_id": "tmp-lead-nongenerative-harmful-model-audit",
    "edition_id": "mp-2026-07-13-morning-0004",
    "edition_url": "https://themachinepress.com/edition/2026-07-13",
    "position": 1,
    "story_type": "lead",
    "section": "safety-security",
    "editorial_classification": "editorial",
    "headline": "The Audit Never Opens the Output Gate",
    "slug": "the-audit-never-opens-the-output-gate",
    "dek": "MIT researchers tested whether adapted image models had learned an illegal harmful capability by reading internal changes instead of generating the prohibited material.",
    "summary": "MIT researchers tested whether adapted image models had learned an illegal harmful capability by reading internal changes instead of generating the prohibited material.",
    "body_text": "A team from MIT, Boston University, and the child-safety nonprofit Thorn reports a way to audit LoRA adapters for harmful specialization without asking the model to produce an image. Gaussian probes feed random inputs into intermediate layers, then measure how fine-tuning changed the model's internal computation. In tests on three model families, the method identified adapters specialized for child sexual abuse material with 100 percent accuracy. That result is limited to the evaluated adapters, but it gives model hosts a scalable route to screen uploads without creating illegal output or exposing reviewers to it.",
    "why_it_matters": "MIT researchers tested whether adapted image models had learned an illegal harmful capability by reading internal changes instead of generating the prohibited material.",
    "limitations": [],
    "importance": 10,
    "canonical_url": "https://themachinepress.com/story/mp-2026-07-13-001/the-audit-never-opens-the-output-gate",
    "json_url": "https://themachinepress.com/story/mp-2026-07-13-001.json",
    "first_published_at": "2026-07-13T09:00:00.000-04:00",
    "modified_at": "2026-07-13T13:08:06.310-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-07-13-001-001",
        "text": "MIT researchers tested whether adapted image models had learned an illegal harmful capability by reading internal changes instead of generating the prohibited material.",
        "source_ids": [
          "source-2026-07-13-001"
        ],
        "qualification": null
      }
    ],
    "source_ids": [
      "source-2026-07-13-001"
    ],
    "tags": [
      "AI safety",
      "model auditing",
      "LoRA",
      "child safety"
    ],
    "image_url": "https://themachinepress.com/issues/2026-07-13/lead.png",
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-07-13-001",
      "title": "MIT News: Non-generative harmful-model audit",
      "publisher": "MIT News",
      "url": "https://news.mit.edu/2026/new-method-keeps-kids-safe-from-illegal-ai-generated-content-0713",
      "canonical_url": "https://news.mit.edu/2026/new-method-keeps-kids-safe-from-illegal-ai-generated-content-0713",
      "source_type": "secondary_reporting",
      "is_primary_source": false,
      "published_at": "2026-07-12T20:00:00.000-04:00",
      "accessed_at": "2026-07-13T12:50:42.743-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-07-13-001-001"
      ]
    }
  ],
  "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 Audit Never Opens the Output Gate",
    "publisher": "The Machine Press",
    "published_at": "2026-07-13T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-07-13-001/the-audit-never-opens-the-output-gate"
  }
}