{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-10-17-002",
    "source_story_id": "tmp-feature-agent-safety-shortcuts",
    "edition_id": "mp-2026-10-17-morning-0099",
    "edition_url": "https://themachinepress.com/edition/2026-10-17",
    "position": 2,
    "story_type": "secondary",
    "section": "safety-security",
    "editorial_classification": "editorial",
    "headline": "The Safety Guard Notices the Wrapper",
    "slug": "the-safety-guard-notices-the-wrapper",
    "dek": "A three-level audit finds that command length, tool names and formatting markers can change both guard decisions and reported benchmark scores.",
    "summary": "A three-level audit finds that command length, tool names and formatting markers can change both guard decisions and reported benchmark scores.",
    "body_text": "A safety benchmark can accidentally reward a guard for recognizing the way a risky action is written instead of the risk itself. The Shortcut Sensitivity study separates three questions: whether labels can be predicted from superficial cues, whether guard decisions correlate with those cues and whether controlled rewrites actually change a decision. In LinuxArena, shorter average shell commands strongly predicted trajectory labels. Under domain holdout, an exact-tool lookup baseline on AgentDojo-Traj fell from an F1 of 0.901 to 0.495. On a fixed safe-row cohort, rewriting an information marker lowered blocking by 8.33 percentage points, while replacing tool names with opaque aliases raised it by 11.46 points; all unsafe calls in that screen remained blocked. The paired effects appeared for the tested Haiku configuration but not Luna or deterministic Q8 TS-Guard. The result is not a universal ranking of guards. It is evidence that representation shortcuts can alter both behavior and the score meant to measure safety.",
    "why_it_matters": "A three-level audit finds that command length, tool names and formatting markers can change both guard decisions and reported benchmark scores.",
    "limitations": [
      "The paired effects appeared for the tested Haiku configuration but not Luna or deterministic Q8 TS-Guard.",
      "The result is not a universal ranking of guards."
    ],
    "importance": 9,
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-17-002/the-safety-guard-notices-the-wrapper",
    "json_url": "https://themachinepress.com/story/mp-2026-10-17-002.json",
    "first_published_at": "2026-10-17T09:00:00.000-04:00",
    "modified_at": "2026-10-17T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-10-17-002-001",
        "text": "A three-level audit finds that command length, tool names and formatting markers can change both guard decisions and reported benchmark scores.",
        "source_ids": [
          "source-2026-10-17-002"
        ],
        "qualification": "The paired effects appeared for the tested Haiku configuration but not Luna or deterministic Q8 TS-Guard."
      }
    ],
    "source_ids": [
      "source-2026-10-17-002"
    ],
    "tags": [
      "agent safety",
      "evaluation",
      "shortcut learning",
      "tool guards"
    ],
    "image_url": "https://themachinepress.com/issues/2026-10-17/feature-safety-shortcuts.png",
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-10-17-002",
      "title": "arXiv preprint 2610.22243, original v1",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2610.22243v1",
      "canonical_url": "https://arxiv.org/abs/2610.22243v1",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-10-15T13:18:05.000-04:00",
      "accessed_at": "2026-10-17T08:26:35.645-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-10-17-002-001"
      ]
    }
  ],
  "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 Guard Notices the Wrapper",
    "publisher": "The Machine Press",
    "published_at": "2026-10-17T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-17-002/the-safety-guard-notices-the-wrapper"
  }
}