{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-10-05-008",
    "source_story_id": "tmp-story-vla-adversarial-patch-defense",
    "edition_id": "mp-2026-10-05-morning-0088",
    "edition_url": "https://themachinepress.com/edition/2026-10-05",
    "position": 8,
    "story_type": "dispatch",
    "section": "safety-security",
    "editorial_classification": "editorial",
    "headline": "The Robot Defense Worked Only When It Knew When to Intervene",
    "slug": "the-robot-defense-worked-only-when-it-knew-when-to-intervene",
    "dek": "A sparse-autoencoder feature triggered selective suppression of adversarial patches in a vision-language-action policy.",
    "summary": "A sparse-autoencoder feature triggered selective suppression of adversarial patches in a vision-language-action policy.",
    "body_text": "Researchers identified an internal VLA feature correlated with adversarial patches, then suppressed it at inference time only when a linear probe detected an attack. On LIBERO-10, conditional intervention improved success under intermittent attacks, while continuously suppressing the same feature substantially harmed ordinary policy performance. The study tests one mechanistic defense setting and does not establish broad patch immunity.",
    "why_it_matters": "A sparse-autoencoder feature triggered selective suppression of adversarial patches in a vision-language-action policy.",
    "limitations": [
      "Researchers identified an internal VLA feature correlated with adversarial patches, then suppressed it at inference time only when a linear probe detected an attack.",
      "The study tests one mechanistic defense setting and does not establish broad patch immunity."
    ],
    "importance": 9,
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-05-008/the-robot-defense-worked-only-when-it-knew-when-to-intervene",
    "json_url": "https://themachinepress.com/story/mp-2026-10-05-008.json",
    "first_published_at": "2026-10-05T09:00:00.000-04:00",
    "modified_at": "2026-10-05T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-10-05-008-001",
        "text": "A sparse-autoencoder feature triggered selective suppression of adversarial patches in a vision-language-action policy.",
        "source_ids": [
          "source-2026-10-05-008"
        ],
        "qualification": "Researchers identified an internal VLA feature correlated with adversarial patches, then suppressed it at inference time only when a linear probe detected an attack."
      }
    ],
    "source_ids": [
      "source-2026-10-05-008"
    ],
    "tags": [
      "VLA",
      "adversarial patches",
      "robot safety"
    ],
    "image_url": "https://themachinepress.com/issues/2026-10-05/vla-defense-robot-arm.webp",
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-10-05-008",
      "title": "arXiv preprint 2610.03498",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2610.03498",
      "canonical_url": "https://arxiv.org/abs/2610.03498",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-10-02T11:57:04.000-04:00",
      "accessed_at": "2026-10-05T08:30:00.000-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-10-05-008-001"
      ]
    }
  ],
  "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 Robot Defense Worked Only When It Knew When to Intervene",
    "publisher": "The Machine Press",
    "published_at": "2026-10-05T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-05-008/the-robot-defense-worked-only-when-it-knew-when-to-intervene"
  }
}