{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-10-23-026",
    "source_story_id": "tmp-story-simulated-control-evals",
    "edition_id": "mp-2026-10-23-morning-0105",
    "edition_url": "https://themachinepress.com/edition/2026-10-23",
    "position": 15,
    "story_type": "dispatch",
    "section": "safety-security",
    "editorial_classification": "editorial",
    "headline": "The Safety Test Simulates the Tools, Not Just the Attacker",
    "slug": "the-safety-test-simulates-the-tools-not-just-the-attacker",
    "dek": "On LinuxArena, simulated control evaluations predicted five monitors' safety with 3.6 percentage points of error at a 5 percent audit budget.",
    "summary": "On LinuxArena, simulated control evaluations predicted five monitors' safety with 3.6 percentage points of error at a 5 percent audit budget.",
    "body_text": "Control evaluations often test an adversarial model in a purpose-built environment, then extrapolate to deployment. Simulated control evaluations replace real tool responses with model-generated ones tailored to the deployment task. Against executable LinuxArena ground truth, they predicted safety across five monitors with 3.6 percentage points of root-mean-square error at a 5 percent audit budget, versus 10.1 points for the separate BashArena setting. They also reproduced attack and monitor behavior more closely and used fewer than 5 percent as many task-pair draws. The result is evidence for the reported environments, not proof that simulation faithfully covers every deployment.",
    "why_it_matters": "On LinuxArena, simulated control evaluations predicted five monitors' safety with 3.6 percentage points of error at a 5 percent audit budget.",
    "limitations": [
      "The result is evidence for the reported environments, not proof that simulation faithfully covers every deployment."
    ],
    "importance": 8,
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-23-026/the-safety-test-simulates-the-tools-not-just-the-attacker",
    "json_url": "https://themachinepress.com/story/mp-2026-10-23-026.json",
    "first_published_at": "2026-10-23T09:00:00.000-04:00",
    "modified_at": "2026-10-23T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-10-23-026-001",
        "text": "On LinuxArena, simulated control evaluations predicted five monitors' safety with 3.6 percentage points of error at a 5 percent audit budget.",
        "source_ids": [
          "source-2026-10-23-015"
        ],
        "qualification": "The result is evidence for the reported environments, not proof that simulation faithfully covers every deployment."
      }
    ],
    "source_ids": [
      "source-2026-10-23-015"
    ],
    "tags": [
      "control evaluations",
      "AI safety",
      "simulation"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-10-23-015",
      "title": "arXiv preprint 2610.31593, original v1",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2610.31593v1",
      "canonical_url": "https://arxiv.org/abs/2610.31593v1",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-10-22T13:26:42.000-04:00",
      "accessed_at": "2026-10-23T08:21:38.024-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-10-23-026-001"
      ]
    }
  ],
  "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 Test Simulates the Tools, Not Just the Attacker",
    "publisher": "The Machine Press",
    "published_at": "2026-10-23T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-23-026/the-safety-test-simulates-the-tools-not-just-the-attacker"
  }
}