{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-08-09-018",
    "source_story_id": "tmp-sidebar-multilingual-rag-privacy",
    "edition_id": "mp-2026-08-09-morning-0031",
    "edition_url": "https://themachinepress.com/edition/2026-08-09",
    "position": 20,
    "story_type": "ticker",
    "section": "safety",
    "editorial_classification": "editorial",
    "headline": "The Leak Lived in the Pipeline, Not the Language Label",
    "slug": "the-leak-lived-in-the-pipeline-not-the-language-label",
    "dek": "English leaked most under one filter, while Arabic and Swahili retained residuals after an added input judge.",
    "summary": "English leaked most under one filter, while Arabic and Swahili retained residuals after an added input judge.",
    "body_text": "A five-language audit of one Qwen-based RAG pipeline found English had the highest unstructured synthetic-PII leak rate with output filtering alone. Adding an input judge left residual leaks in Arabic and Swahili; the authors explicitly frame the result as pipeline-conditional and do not measure benign-query false positives or utility cost.",
    "why_it_matters": "English leaked most under one filter, while Arabic and Swahili retained residuals after an added input judge.",
    "limitations": [
      "Adding an input judge left residual leaks in Arabic and Swahili; the authors explicitly frame the result as pipeline-conditional and do not measure benign-query false positives or utility cost."
    ],
    "importance": 7,
    "canonical_url": "https://themachinepress.com/story/mp-2026-08-09-018/the-leak-lived-in-the-pipeline-not-the-language-label",
    "json_url": "https://themachinepress.com/story/mp-2026-08-09-018.json",
    "first_published_at": "2026-08-09T09:00:00.000-04:00",
    "modified_at": "2026-08-09T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-08-09-018-001",
        "text": "English leaked most under one filter, while Arabic and Swahili retained residuals after an added input judge.",
        "source_ids": [
          "source-2026-08-09-020"
        ],
        "qualification": "Adding an input judge left residual leaks in Arabic and Swahili; the authors explicitly frame the result as pipeline-conditional and do not measure benign-query false positives or utility cost."
      }
    ],
    "source_ids": [
      "source-2026-08-09-020"
    ],
    "tags": [
      "RAG",
      "privacy",
      "multilingual AI"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-08-09-020",
      "title": "arXiv preprint 2608.05163",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2608.05163",
      "canonical_url": "https://arxiv.org/abs/2608.05163",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": null,
      "accessed_at": "2026-08-09T08:17:45.000-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-08-09-018-001"
      ]
    }
  ],
  "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 Leak Lived in the Pipeline, Not the Language Label",
    "publisher": "The Machine Press",
    "published_at": "2026-08-09T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-08-09-018/the-leak-lived-in-the-pipeline-not-the-language-label"
  }
}