{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-09-27-007",
    "source_story_id": "tmp-story-endoprompt-utility-degradation",
    "edition_id": "mp-2026-09-27-morning-0080",
    "edition_url": "https://themachinepress.com/edition/2026-09-27",
    "position": 7,
    "story_type": "dispatch",
    "section": "safety-security",
    "editorial_classification": "editorial",
    "headline": "A Learned Prefix Cut Model Utility by 26.8 Points",
    "slug": "a-learned-prefix-cut-model-utility-by-26-8-points",
    "dek": "ENDOPROMPT optimized harmful prefixes from unlabeled instructions and produced losses across almost every tested split.",
    "summary": "ENDOPROMPT optimized harmful prefixes from unlabeled instructions and produced losses across almost every tested split.",
    "body_text": "ENDOPROMPT uses white-box optimization to derive input prefixes without requiring labeled attack examples. Across four models, seven benign benchmarks and 28 model-benchmark combinations, the authors report a mean utility loss of 26.8 percentage points and negative effects in 27 of 28 cases. Their controls did not establish a compensating request-matching benefit; the preprint promises code upon acceptance, so independent reproduction remains a next step.",
    "why_it_matters": "ENDOPROMPT optimized harmful prefixes from unlabeled instructions and produced losses across almost every tested split.",
    "limitations": [
      "Their controls did not establish a compensating request-matching benefit; the preprint promises code upon acceptance, so independent reproduction remains a next step."
    ],
    "importance": 9,
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-27-007/a-learned-prefix-cut-model-utility-by-26-8-points",
    "json_url": "https://themachinepress.com/story/mp-2026-09-27-007.json",
    "first_published_at": "2026-09-27T09:00:00.000-04:00",
    "modified_at": "2026-09-27T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-09-27-007-001",
        "text": "ENDOPROMPT optimized harmful prefixes from unlabeled instructions and produced losses across almost every tested split.",
        "source_ids": [
          "source-2026-09-27-007"
        ],
        "qualification": "Their controls did not establish a compensating request-matching benefit; the preprint promises code upon acceptance, so independent reproduction remains a next step."
      }
    ],
    "source_ids": [
      "source-2026-09-27-007"
    ],
    "tags": [
      "prompt attacks",
      "model utility",
      "adversarial optimization"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-09-27-007",
      "title": "arXiv preprint 2609.29948",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2609.29948",
      "canonical_url": "https://arxiv.org/abs/2609.29948",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-09-24T11:08:13.000-04:00",
      "accessed_at": "2026-09-27T08:26:41.000-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-09-27-007-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "A Learned Prefix Cut Model Utility by 26.8 Points",
    "publisher": "The Machine Press",
    "published_at": "2026-09-27T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-27-007/a-learned-prefix-cut-model-utility-by-26-8-points"
  }
}