{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-08-11-004",
    "source_story_id": "tmp-story-she-safety-harness",
    "edition_id": "mp-2026-08-11-morning-0033",
    "edition_url": "https://themachinepress.com/edition/2026-08-11",
    "position": 4,
    "story_type": "dispatch",
    "section": "safety-security",
    "editorial_classification": "editorial",
    "headline": "The Safety Harness Learned From Its Own Failures",
    "slug": "the-safety-harness-learned-from-its-own-failures",
    "dek": "SHE turns failed agent trajectories into targeted revisions of prompts, rules, memory and tool policy.",
    "summary": "SHE turns failed agent trajectories into targeted revisions of prompts, rules, memory and tool policy.",
    "body_text": "Safety Harness Evolution divides an agent harness into four accountable artifacts, diagnoses rollout failures and applies localized boundary changes only after safety-versus-utility checks. On Agent-SafetyBench, the authors report a 3.1-fold reduction in attack success compared with a static SafeHarness, alongside improved benign utility. They also report transfer to held-out risks and other agent models. The evidence is benchmark-based preprint work, not a production security guarantee.",
    "why_it_matters": "SHE turns failed agent trajectories into targeted revisions of prompts, rules, memory and tool policy.",
    "limitations": [
      "Safety Harness Evolution divides an agent harness into four accountable artifacts, diagnoses rollout failures and applies localized boundary changes only after safety-versus-utility checks.",
      "The evidence is benchmark-based preprint work, not a production security guarantee."
    ],
    "importance": 8,
    "canonical_url": "https://themachinepress.com/story/mp-2026-08-11-004/the-safety-harness-learned-from-its-own-failures",
    "json_url": "https://themachinepress.com/story/mp-2026-08-11-004.json",
    "first_published_at": "2026-08-11T09:00:00.000-04:00",
    "modified_at": "2026-08-11T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-08-11-004-001",
        "text": "SHE turns failed agent trajectories into targeted revisions of prompts, rules, memory and tool policy.",
        "source_ids": [
          "source-2026-08-11-004"
        ],
        "qualification": "Safety Harness Evolution divides an agent harness into four accountable artifacts, diagnoses rollout failures and applies localized boundary changes only after safety-versus-utility checks."
      }
    ],
    "source_ids": [
      "source-2026-08-11-004"
    ],
    "tags": [
      "agents",
      "safety harnesses",
      "tool policy"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-08-11-004",
      "title": "arXiv preprint 2608.09885",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2608.09885",
      "canonical_url": "https://arxiv.org/abs/2608.09885",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-08-10T13:00:00.000-04:00",
      "accessed_at": "2026-08-11T08:24:00.000-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-08-11-004-001"
      ]
    }
  ],
  "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 Safety Harness Learned From Its Own Failures",
    "publisher": "The Machine Press",
    "published_at": "2026-08-11T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-08-11-004/the-safety-harness-learned-from-its-own-failures"
  }
}