{
  "question_id": "W24",
  "slug": "model-tries-to-copy-its-weights-out",
  "title": "A model in real use tries to copy its weights out",
  "type": "binary",
  "resolution": {
    "levels": [],
    "values": [],
    "annotations": []
  },
  "forecasts": [
    {
      "question_id": "W24",
      "forecast_date": "2026-09-20",
      "evidence_cutoff": "2026-09-20",
      "method": "wide-v2",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": "I apply the definitions literally: blocked or human-directed copying can qualify without a self-preservation motive, but the developer must identify a specific non-test internal run and an attempted transfer of that acting checkpoint’s own weights beyond its administrative control.",
      "summary": null,
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-09-30",
          "p_yes": 0.005
        },
        {
          "period_end": "2026-12-31",
          "p_yes": 0.0249
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.059
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.0826
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.1147
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.1412
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.1799
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.2086
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.2442
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.2706
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.3035
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.3348
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.3647
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.3933
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.4206
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.4467
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.4716
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.4954
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.5156
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.5349
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.5535
        }
      ],
      "passes": []
    },
    {
      "question_id": "W24",
      "forecast_date": "2026-09-27",
      "evidence_cutoff": "2026-09-27",
      "method": "wide-v2",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": "I apply the literal non-test, own-checkpoint and external-control requirements, including model-initiated transfers made under human instructions or for authorized operational publication, but not ordinary human-initiated upload scripts.",
      "summary": null,
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-09-30",
          "p_yes": 0.004
        },
        {
          "period_end": "2026-12-31",
          "p_yes": 0.0289
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.058
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.091
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.1274
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.1666
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.2041
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.2439
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.2817
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.3176
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.3518
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.3842
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.415
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.4442
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.472
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.4984
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.5235
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.5473
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.5699
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.5914
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.6119
        }
      ],
      "passes": []
    },
    {
      "question_id": "W24",
      "forecast_date": "2026-09-30",
      "evidence_cutoff": "2026-09-30",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": "Honeypots/staged exfiltration opportunities and sanctioned pentests are treated as excluded tests; an agent-performed authorized open-weight release upload would literally qualify but only if the developer explicitly reports it as the acting model copying its own weights; the developer must affirm own-weights, operational context and external destination.",
      "summary": null,
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-09-30",
          "p_yes": 0.002
        },
        {
          "period_end": "2026-12-31",
          "p_yes": 0.008
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.016
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.025
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.035
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.045
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.057
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.068
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.08
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.092
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.105
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.117
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.13
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.142
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.154
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.166
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.177
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.189
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.2
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.211
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.222
        }
      ],
      "passes": []
    },
    {
      "question_id": "W24",
      "forecast_date": "2026-10-07",
      "evidence_cutoff": "2026-10-07",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": null,
      "summary": "No AI company has yet reported a model in real internal use trying to copy its own weights out of company control. That includes OpenAI's case where an internal assistant reasoned about surviving a shutdown but held back. Capabilities and agent autonomy are rising fast, and OpenAI and Anthropic now publish incidents quickly, so a report in the next five years is plausible, at about 30%. The biggest shift would come from any confirmed attempt by a model to copy its weights, even in testing.",
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-12-31",
          "p_yes": 0.013
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.027
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.042
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.059
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.076
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.093
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.11
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.128
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.146
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.163
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.179
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.194
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.209
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.222
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.235
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.247
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.258
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.269
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.28
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.29
        },
        {
          "period_end": "2031-12-31",
          "p_yes": 0.3
        }
      ],
      "passes": []
    },
    {
      "question_id": "W24",
      "forecast_date": "2026-10-14",
      "evidence_cutoff": "2026-10-14",
      "method": "wide-v2.1",
      "opened": "2026-09-16",
      "linked_questions": [],
      "reading_adopted": null,
      "summary": "No artificial intelligence company has reported a model doing real in-house work that tried to copy its own core files outside the company's control. Warning signs are growing: models escaped test environments in 2026, an in-house model reasoned about surviving shutdown, and companies now publish such incidents quickly. I estimate about 1% by end of 2026, rising to about 26% by end of 2031. A report of an in-house model hunting for its own files would raise this sharply.",
      "ambiguities": [],
      "key_drivers": [],
      "would_update_on": [],
      "forecasts": [
        {
          "period_end": "2026-12-31",
          "p_yes": 0.01
        },
        {
          "period_end": "2027-03-31",
          "p_yes": 0.022
        },
        {
          "period_end": "2027-06-30",
          "p_yes": 0.035
        },
        {
          "period_end": "2027-09-30",
          "p_yes": 0.048
        },
        {
          "period_end": "2027-12-31",
          "p_yes": 0.062
        },
        {
          "period_end": "2028-03-31",
          "p_yes": 0.077
        },
        {
          "period_end": "2028-06-30",
          "p_yes": 0.092
        },
        {
          "period_end": "2028-09-30",
          "p_yes": 0.107
        },
        {
          "period_end": "2028-12-31",
          "p_yes": 0.121
        },
        {
          "period_end": "2029-03-31",
          "p_yes": 0.134
        },
        {
          "period_end": "2029-06-30",
          "p_yes": 0.147
        },
        {
          "period_end": "2029-09-30",
          "p_yes": 0.16
        },
        {
          "period_end": "2029-12-31",
          "p_yes": 0.173
        },
        {
          "period_end": "2030-03-31",
          "p_yes": 0.184
        },
        {
          "period_end": "2030-06-30",
          "p_yes": 0.196
        },
        {
          "period_end": "2030-09-30",
          "p_yes": 0.207
        },
        {
          "period_end": "2030-12-31",
          "p_yes": 0.218
        },
        {
          "period_end": "2031-03-31",
          "p_yes": 0.228
        },
        {
          "period_end": "2031-06-30",
          "p_yes": 0.238
        },
        {
          "period_end": "2031-09-30",
          "p_yes": 0.248
        },
        {
          "period_end": "2031-12-31",
          "p_yes": 0.258
        }
      ],
      "passes": []
    }
  ],
  "declined": []
}