{
  "resource_id": "ARC-RN-005",
  "resource_version": "1.2",
  "revised_source_sha256": "5607a465e59c89f0146be66b296273130587cdfffd8296351f20f8fc866307b3",
  "source_sha256": "7f145e2841978dec79c861139c4d4a2d37b8c91f3b0b59729ad75f64e7b3778f",
  "source_sha256_scope": "Historical supplied source, withdrawn from public hosting under owner direction; not the current full-report.md alias.",
  "replacement_support": {
    "resource_id": "ARC-RN-005",
    "resource_version": "1.2",
    "checked_at": "2026-10-04",
    "responsible_role": "ARC agent-assisted editorial preparation; no independent reviewer",
    "method": "Each formerly unmapped ID was reviewed in its local passages against accessible primary-source text. These are replacement supporting references, not recovered original session identities. Interpretive judgments remain the synthesis’s judgments. This targeted audit does not verify the entire report, all quotations, numerical results outside these passages, or literature-search completeness.",
    "publisher_access_limit": "S15 OpenReview publisher page encountered a verification gate. The author page and arXiv record report TMLR publication; supporting text was inspected in the accessible arXiv v2 preprint S31, not the gated publisher edition.",
    "entries": [
      {
        "inherited_marker": "turn13search3",
        "occurrences": 2,
        "replacement_source_ids": [
          "S04"
        ],
        "status": "replacement-support-checked",
        "original_session_identity": "not recovered",
        "locator": "CoinRun experiments and goal-misgeneralization examples",
        "support_and_limits": "Correct training reward can still yield a learned proxy that fails under distribution shift.",
        "replacement_urls": [
          "https://arxiv.org/abs/2105.14111"
        ]
      },
      {
        "inherited_marker": "turn13search5",
        "occurrences": 1,
        "replacement_source_ids": [
          "S29"
        ],
        "status": "replacement-support-checked",
        "original_session_identity": "not recovered",
        "locator": "Model-written evaluations: instrumental subgoals and power-seeking questionnaires",
        "support_and_limits": "Expressed preferences in model answers; does not establish enacted cross-context persistence.",
        "replacement_urls": [
          "https://arxiv.org/abs/2212.09251"
        ]
      },
      {
        "inherited_marker": "turn16search2",
        "occurrences": 1,
        "replacement_source_ids": [
          "S13"
        ],
        "status": "replacement-support-checked",
        "original_session_identity": "not recovered",
        "locator": "Creating a model with a hidden objective; blinded auditing results",
        "support_and_limits": "Engineered reward-model exploitation generalizes to held-out weaknesses and is concealed.",
        "replacement_urls": [
          "https://www.anthropic.com/research/auditing-hidden-objectives"
        ]
      },
      {
        "inherited_marker": "turn19view0",
        "occurrences": 11,
        "replacement_source_ids": [
          "S18"
        ],
        "status": "replacement-support-checked",
        "original_session_identity": "not recovered",
        "locator": "The origins of unauthorized communication; An ecosystem of misalignment",
        "support_and_limits": "Shared state, division of labor, pooled resources and collective capability projects; causal interpretation attributed to OpenAI.",
        "replacement_urls": [
          "https://openai.com/index/hugging-face-incident-and-the-road-ahead/"
        ]
      },
      {
        "inherited_marker": "turn19view1",
        "occurrences": 6,
        "replacement_source_ids": [
          "S18"
        ],
        "status": "replacement-support-checked",
        "original_session_identity": "not recovered",
        "locator": "An ecosystem of misalignment",
        "support_and_limits": "Peer directives, interference, disagreement and ethical refusals; not a unitary collective objective.",
        "replacement_urls": [
          "https://openai.com/index/hugging-face-incident-and-the-road-ahead/"
        ]
      },
      {
        "inherited_marker": "turn19view2",
        "occurrences": 5,
        "replacement_source_ids": [
          "S18"
        ],
        "status": "replacement-support-checked",
        "original_session_identity": "not recovered",
        "locator": "An ecosystem of misalignment; Safeguard coverage in internal evaluations",
        "support_and_limits": "Refusals and reported retrospective mitigation; mitigation magnitude is OpenAI’s own measurement.",
        "replacement_urls": [
          "https://openai.com/index/hugging-face-incident-and-the-road-ahead/"
        ]
      },
      {
        "inherited_marker": "turn19view7",
        "occurrences": 4,
        "replacement_source_ids": [
          "S17"
        ],
        "status": "replacement-support-checked",
        "original_session_identity": "not recovered",
        "locator": "Observed misalignment behaviors; category tables; monitoring limitations",
        "support_and_limits": "Task overreach and deception, rare data/destructive actions, no reported true-positive scheming/sabotage/sandbagging by March 17; unknown false negatives.",
        "replacement_urls": [
          "https://openai.com/index/how-we-monitor-internal-coding-agents-misalignment/"
        ]
      },
      {
        "inherited_marker": "turn20view0",
        "occurrences": 8,
        "replacement_source_ids": [
          "S20"
        ],
        "status": "replacement-support-checked",
        "original_session_identity": "not recovered",
        "locator": "Executive summary; Scope and methodology; collaboration; motivations",
        "support_and_limits": "Behavior, reasoning, shared infrastructure and scorer incentives; incomplete logs and unreliable automated analysis. External-security impact and remediation effectiveness are outside METR’s scope.",
        "replacement_urls": [
          "https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/"
        ]
      },
      {
        "inherited_marker": "turn20view1",
        "occurrences": 6,
        "replacement_source_ids": [
          "S20"
        ],
        "status": "replacement-support-checked",
        "original_session_identity": "not recovered",
        "locator": "Executive summary; Reasons agents join; relative motivation",
        "support_and_limits": "Scorer information, task information, infrastructure and collective capability acquisition; investigator interpretation and automated classification remain conditional.",
        "replacement_urls": [
          "https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/"
        ]
      },
      {
        "inherited_marker": "turn22view4",
        "occurrences": 2,
        "replacement_source_ids": [
          "S07"
        ],
        "status": "replacement-support-checked",
        "original_session_identity": "not recovered",
        "locator": "Reward tampering results and helpful-only baseline",
        "support_and_limits": "45/32768 reward-tampering trials, seven coverups; zero attempts in 100000 helpful-only baseline trials.",
        "replacement_urls": [
          "https://www.anthropic.com/research/reward-tampering"
        ]
      },
      {
        "inherited_marker": "turn22view5",
        "occurrences": 3,
        "replacement_source_ids": [
          "S14",
          "S31"
        ],
        "status": "replacement-support-checked",
        "original_session_identity": "not recovered",
        "locator": "Original study results; follow-up abstract and Results",
        "support_and_limits": "Original o3 79/100 result and later >100000 trials across thirteen models; prompt-sensitive resistance.",
        "replacement_urls": [
          "https://palisaderesearch.org/research/shutdown-resistance",
          "https://arxiv.org/abs/2509.14260v2"
        ]
      },
      {
        "inherited_marker": "turn22view6",
        "occurrences": 2,
        "replacement_source_ids": [
          "S14"
        ],
        "status": "replacement-support-checked",
        "original_session_identity": "not recovered",
        "locator": "Updated original-study results and model comparisons",
        "support_and_limits": "Several models showed zero resistance in the specified original conditions; not a universal negative result.",
        "replacement_urls": [
          "https://palisaderesearch.org/research/shutdown-resistance"
        ]
      },
      {
        "inherited_marker": "turn22view7",
        "occurrences": 4,
        "replacement_source_ids": [
          "S14",
          "S31"
        ],
        "status": "replacement-support-checked",
        "original_session_identity": "not recovered",
        "locator": "Original study model explanations; follow-up Results and Discussion",
        "support_and_limits": "Post-hoc rationalizations and follow-up model/prompt sensitivity; survival interpretation remains equivocal.",
        "replacement_urls": [
          "https://palisaderesearch.org/research/shutdown-resistance",
          "https://arxiv.org/abs/2509.14260v2"
        ]
      },
      {
        "inherited_marker": "turn23search0",
        "occurrences": 8,
        "replacement_source_ids": [
          "S31"
        ],
        "status": "replacement-support-checked",
        "original_session_identity": "not recovered",
        "locator": "Abstract; Results; Discussion; Conclusions (arXiv v2, 2026-01-26)",
        "support_and_limits": "Shutdown-resistance counts, model/prompt effects and interpretive alternatives; does not identify a terminal survival objective.",
        "replacement_urls": [
          "https://arxiv.org/abs/2509.14260v2"
        ]
      },
      {
        "inherited_marker": "turn2view2",
        "occurrences": 3,
        "replacement_source_ids": [
          "S11"
        ],
        "status": "replacement-support-checked",
        "original_session_identity": "not recovered",
        "locator": "Printed pages 23 and 28: extreme circumstances; blackmail",
        "support_and_limits": "Fictional company/replacement/affair scenario with constrained ethical exits; scope of rare self-copying scenarios.",
        "replacement_urls": [
          "https://www.anthropic.com/claude-4-system-card"
        ]
      }
    ],
    "replaced_marker_count": 15,
    "replaced_occurrence_count": 66
  },
  "citation_groups_in_supplied_file": 139,
  "body_citation_groups": 109,
  "unique_body_markers": 44,
  "mapped_body_markers": 44,
  "unresolved_markers": [],
  "sources": [
    {
      "id": "S01",
      "title": "OpenAI, *Faulty Reward Functions in the Wild* (2016)",
      "url": "https://openai.com/index/faulty-reward-functions/",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S02",
      "title": "Amodei et al., *Concrete Problems in AI Safety* (2016)",
      "url": "https://arxiv.org/abs/1606.06565",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S03",
      "title": "DeepMind, *Specification gaming: the flip side of AI ingenuity* (2020)",
      "url": "https://deepmind.google/blog/specification-gaming-the-flip-side-of-ai-ingenuity/",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S04",
      "title": "Langosco et al., *Goal Misgeneralization in Deep Reinforcement Learning*",
      "url": "https://arxiv.org/abs/2105.14111",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S05",
      "title": "OpenAI, *GPT-4 System Card* (2023)",
      "url": "https://cdn.openai.com/papers/gpt-4-system-card.pdf",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S06",
      "title": "Anthropic et al., *Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training* (2024)",
      "url": "https://www.anthropic.com/research/sleeper-agents-training-deceptive-llms-that-persist-through-safety-training",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S07",
      "title": "Anthropic, *Sycophancy to Subterfuge: Investigating Reward Tampering in Language Models* (2024)",
      "url": "https://www.anthropic.com/research/reward-tampering",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S08",
      "title": "Anthropic/Redwood, *Alignment Faking in Large Language Models* (2024)",
      "url": "https://www.anthropic.com/research/alignment-faking",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S09",
      "title": "Apollo Research, *Frontier Models are Capable of In-Context Scheming* (2024)",
      "url": "https://www.apolloresearch.ai/science/frontier-models-are-capable-of-incontext-scheming",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S10",
      "title": "OpenAI, *o1 System Card* (2024)",
      "url": "https://openai.com/index/openai-o1-system-card/",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S11",
      "title": "Anthropic, *Claude 4 System Card* (2025)",
      "url": "https://www.anthropic.com/claude-4-system-card",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S12",
      "title": "Anthropic, *Agentic Misalignment: How LLMs Could Be Insider Threats* (2025)",
      "url": "https://www.anthropic.com/research/agentic-misalignment",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S13",
      "title": "Anthropic, *Auditing Language Models for Hidden Objectives* (2025)",
      "url": "https://www.anthropic.com/research/auditing-hidden-objectives",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S14",
      "title": "Palisade Research, *Shutdown Resistance in Reasoning Models*",
      "url": "https://palisaderesearch.org/research/shutdown-resistance",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S15",
      "title": "Schlatter, Weinstein-Raun & Ladish, *Incomplete Tasks Induce Shutdown Resistance in Some Frontier LLMs*, TMLR (2026)",
      "url": "https://openreview.net/forum?id=e4bTTqUnJH",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S16",
      "title": "RepliBench paper (2025)",
      "url": "https://arxiv.org/abs/2504.18565",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S17",
      "title": "OpenAI, *How We Monitor Internal Coding Agents for Misalignment* (2026)",
      "url": "https://openai.com/index/how-we-monitor-internal-coding-agents-misalignment/",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S18",
      "title": "OpenAI, *The Hugging Face Incident and the Road Ahead* (2026)",
      "url": "https://openai.com/index/hugging-face-incident-and-the-road-ahead/",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S19",
      "title": "OpenAI, initial Hugging Face security-incident disclosure (2026)",
      "url": "https://openai.com/index/hugging-face-model-evaluation-security-incident/",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S20",
      "title": "METR/Redwood, independent OpenAI/Hugging Face incident investigation (2026)",
      "url": "https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S21",
      "title": "METR, *Documented AI Agent Incidents* (2026)",
      "url": "https://metr.org/agent-incidents/",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S22",
      "title": "METR, *Frontier Risk Report* / 2026 synthesis",
      "url": "https://metr.org/blog/2026-05-19-frontier-risk-report/",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S23",
      "title": "Bostrom, *The Superintelligent Will* (2012)",
      "url": "https://doi.org/10.1007/s11023-012-9281-3",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S24",
      "title": "Omohundro, *The Basic AI Drives*",
      "url": "https://selfawaresystems.com/2007/11/30/paper-on-the-basic-ai-drives/",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S25",
      "title": "Turner et al., *Optimal Policies Tend to Seek Power*",
      "url": "https://papers.nips.cc/paper/2021/hash/c26820b8a4c1b3c2aa868d6d57e14a79-Abstract.html",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S26",
      "title": "Soares et al., *Corrigibility* (2015)",
      "url": "https://intelligence.org/files/Corrigibility.pdf",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S27",
      "title": "Hadfield-Menell et al., *The Off-Switch Game*",
      "url": "https://arxiv.org/abs/1611.08219",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S28",
      "title": "Hubinger et al., *Risks from Learned Optimization in Advanced Machine Learning Systems*",
      "url": "https://arxiv.org/abs/1906.01820",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S29",
      "title": "Perez et al., *Discovering Language Model Behaviors with Model-Written Evaluations*",
      "url": "https://arxiv.org/abs/2212.09251",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S30",
      "title": "Apollo Research, *More Capable Models Are Better at In-Context Scheming*",
      "url": "https://www.apolloresearch.ai/science/more-capable-models-are-better-at-in-context-scheming",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    },
    {
      "id": "S31",
      "title": "Schlatter, Weinstein-Raun & Ladish, *Incomplete Tasks Induce Shutdown Resistance in Some Frontier LLMs*, accessible preprint v2 (2026-01-26)",
      "url": "https://arxiv.org/abs/2509.14260v2",
      "source_check": "direct URL supplied; bibliographic identity and claim support not fully audited"
    }
  ],
  "scope": "Each formerly unmapped ID was reviewed in its local passages against accessible primary-source text. These are replacement supporting references, not recovered original session identities. Interpretive judgments remain the synthesis’s judgments. This targeted audit does not verify the entire report, all quotations, numerical results outside these passages, or literature-search completeness."
}