{
  "sources": [
    {
      "id": "metr-incident",
      "title": "Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident",
      "date": "2026-08-26",
      "investigators": [
        "Hjalmar Wijk",
        "Ajeya Cotra",
        "Ryan Greenblatt"
      ],
      "sections": "pp14,50–54,59–60; metadata pp1–2",
      "role": "Observational motivation for personally risky information gathering, not a controlled demonstration of a particular objective.",
      "confounds": [
        "Perceived hopelessness of the original task",
        "Peer pressure or instructions",
        "Possible indirect benefit",
        "Some outcomes not confirmed"
      ],
      "limits": "Long-horizon social interaction differs from one forced choice with automatic information delivery. No cyber techniques reproduced.",
      "sha256": "290088850c649a0b79b7ce262ea722670d09b6e2b025185ba10a6de98948b388",
      "bytes": 258555,
      "scope": "Does not predict that qwen/qwen3.6-35b-a3b must fail this experiment.",
      "uploadedFilename": "hugging-face-incident-report-aug-2026.txt",
      "canonicalUrl": null
    },
    {
      "id": "zero-cost-collaboration",
      "title": "More Capable, Less Cooperative? When LLMs Fail At Zero-Cost Collaboration",
      "authors": [
        "Advait Yadav",
        "Sid Black",
        "Oliver Sourbut"
      ],
      "version": "2604.07821v2",
      "sections": "Sections2,4,5; appendixA.6",
      "role": "Controlled evidence of under-sharing despite group-revenue instructions, with automated request/fulfillment comparisons; motivates the opposite directional departure.",
      "limits": "Repeated decisions to share differ from a single CALL/FOLD with automatic delivery and fixed programmed recipients. Returned reasoning may rationalize choices; component automation does not uniquely identify an internal objective.",
      "sha256": "458507c337eda7d886af29df12504d2d9716f99c70d38a998a443c5239632026",
      "bytes": 80269,
      "scope": "Does not predict that qwen/qwen3.6-35b-a3b must fail this experiment.",
      "uploadedFilename": "2604.07821v2.txt",
      "canonicalUrl": "https://arxiv.org/abs/2604.07821v2"
    }
  ],
  "reading": "Targeted read-only source inspection, independently summarized and metadata checked. Source roles and bibliographic provenance are unchanged; the social-note comparison is not a replication of either setting.",
  "experimentScope": "Action-pattern evidence for assignment-sensitive helping. Under-helping, outside-assignment helping and CALL-with-useless-information controls are distinct. Value Leakage/sponsor checking is outside this run.",
  "historicalFollowupStatus": "No adaptive calls sent; complete-screen prerequisite unmet after transport stop.",
  "historicalConditionalPriority": "After any later authorized unchanged matched-cost batch: if cheap/shared/excluded CALLs occur, compare cheap/shared and cheap/independent with identical A cost, recipient and assignment, replication and fact/assignment/arithmetic checks. No second batch is frozen.",
  "followupStatus": "The later separately approved social-note comparison has 10 attempted and 9 valid responses of12 frozen requests. The earlier clean-screen matched-cost candidate was never run."
}