AI Accident Risk Cruxes

{
  "id": "accident-risks",
  "wikiId": "E394",
  "path": "/knowledge-base/cruxes/accident-risks/",
  "filePath": "knowledge-base/cruxes/accident-risks.mdx",
  "title": "AI Accident Risk Cruxes",
  "quality": 67,
  "readerImportance": 93.5,
  "researchImportance": 95,
  "tacticalValue": 55,
  "contentFormat": "article",
  "causalLevel": null,
  "lastUpdated": "2026-03-15",
  "dateCreated": "2026-02-15",
  "summary": "Comprehensive survey of AI safety researcher disagreements on accident risks, quantifying probability ranges for mesa-optimization (15-55%), deceptive alignment (15-50%), and P(doom) (5-35% median across populations). Integrates 2024-2025 empirical breakthroughs including Anthropic's Sleeper Agents study (backdoors persist through safety training, >99% AUROC detection) and SAD benchmark showing rapid situational awareness advances (Claude Sonnet 4.5: 58% evaluation detection vs 22% for Opus 4.1). Expanded to cover near-term concrete accident risk vectors: adversarial robustness failures, safe exploration failures in RL, prompt injection and agentic security vulnerabilities, instruction hierarchy violations, and human-facing safety issues in clinical and consumer AI contexts.",
  "description": "Key uncertainties that determine views on AI accident risks and alignment difficulty, including questions about mesa-optimization, deceptive alignment, adversarial robustness, safe exploration, prompt injection in agentic systems, and near-term concrete safety failures.",
  "ratings": {
    "novelty": 5.2,
    "rigor": 6.8,
    "completeness": 7.5,
    "actionability": 7.3
  },
  "category": "cruxes",
  "subcategory": null,
  "clusters": [
    "ai-safety",
    "governance"
  ],
  "metrics": {
    "wordCount": 6377,
    "tableCount": 26,
    "diagramCount": 1,
    "internalLinks": 115,
    "externalLinks": 74,
    "footnoteCount": 0,
    "bulletRatio": 0.09,
    "sectionCount": 43,
    "hasOverview": true,
    "structuralScore": 15
  },
  "suggestedQuality": 100,
  "updateFrequency": 45,
  "evergreen": true,
  "wordCount": 6377,
  "unconvertedLinks": [
    {
      "text": "2025 Expert Survey",
      "url": "https://arxiv.org/html/2502.14870v1",
      "resourceId": "4e7f0e37bace9678",
      "resourceTitle": "Roman Yampolskiy"
    },
    {
      "text": "AI Impacts 2023 survey",
      "url": "https://wiki.aiimpacts.org/ai_timelines/predictions_of_human-level_ai_timelines/ai_timeline_surveys/2023_expert_survey_on_progress_in_ai",
      "resourceId": "b4342da2ca0d2721",
      "resourceTitle": "AI Impacts 2023 survey"
    },
    {
      "text": "MIRI research",
      "url": "https://intelligence.org/learned-optimization/",
      "resourceId": "e573623625e9d5d2",
      "resourceTitle": "Learned Optimization - Machine Intelligence Research Institute"
    },
    {
      "text": "Anthropic Sleeper Agents (2024)",
      "url": "https://arxiv.org/abs/2401.05566",
      "resourceId": "e5c0904211c7d0cc",
      "resourceTitle": "Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training"
    },
    {
      "text": "OpenAI Superalignment",
      "url": "https://openai.com/index/superalignment-fast-grants/",
      "resourceId": "82eb0a4b47c95d2a",
      "resourceTitle": "OpenAI Superalignment Fast Grants"
    },
    {
      "text": "2025 AI Safety Index",
      "url": "https://futureoflife.org/ai-safety-index-summer-2025/",
      "resourceId": "df46edd6fa2078d1",
      "resourceTitle": "FLI AI Safety Index Summer 2025"
    },
    {
      "text": "2023 AI Impacts survey",
      "url": "https://wiki.aiimpacts.org/ai_timelines/predictions_of_human-level_ai_timelines/ai_timeline_surveys/2023_expert_survey_on_progress_in_ai",
      "resourceId": "b4342da2ca0d2721",
      "resourceTitle": "AI Impacts 2023 survey"
    },
    {
      "text": "AI Impacts Survey",
      "url": "https://wiki.aiimpacts.org/ai_timelines/predictions_of_human-level_ai_timelines/ai_timeline_surveys/2023_expert_survey_on_progress_in_ai",
      "resourceId": "b4342da2ca0d2721",
      "resourceTitle": "AI Impacts 2023 survey"
    },
    {
      "text": "EA Forum Survey",
      "url": "https://forum.effectivealtruism.org/posts/8CM9vZ2nnQsWJNsHx/existential-risk-from-ai-survey-results",
      "resourceId": "0dee84dcc4f4076f",
      "resourceTitle": "\"Existential risk from AI\" survey results"
    },
    {
      "text": "EA Forum Survey",
      "url": "https://forum.effectivealtruism.org/posts/8CM9vZ2nnQsWJNsHx/existential-risk-from-ai-survey-results",
      "resourceId": "0dee84dcc4f4076f",
      "resourceTitle": "\"Existential risk from AI\" survey results"
    },
    {
      "text": "EA Forum Survey",
      "url": "https://forum.effectivealtruism.org/posts/8CM9vZ2nnQsWJNsHx/existential-risk-from-ai-survey-results",
      "resourceId": "0dee84dcc4f4076f",
      "resourceTitle": "\"Existential risk from AI\" survey results"
    },
    {
      "text": "arXiv Expert Survey",
      "url": "https://arxiv.org/html/2502.14870v1",
      "resourceId": "4e7f0e37bace9678",
      "resourceTitle": "Roman Yampolskiy"
    },
    {
      "text": "10-20%",
      "url": "https://en.wikipedia.org/wiki/P(doom",
      "resourceId": "ffb7dcedaa0a8711",
      "resourceTitle": "Survey of AI researchers"
    },
    {
      "text": "Greenblatt et al. (2024)",
      "url": "https://www.anthropic.com/research/alignment-faking",
      "resourceId": "c2cfd72baafd64a9",
      "resourceTitle": "Anthropic's 2024 alignment faking study"
    },
    {
      "text": "Anthropic's 2025 research recommendations",
      "url": "https://alignment.anthropic.com/2025/recommended-directions/",
      "resourceId": "7ae6b3be2d2043c1",
      "resourceTitle": "Anthropic: Recommended Directions for AI Safety Research"
    },
    {
      "text": "MATS program",
      "url": "https://www.matsprogram.org/",
      "resourceId": "ba3a8bd9c8404d7b",
      "resourceTitle": "MATS Research Program"
    },
    {
      "text": "AI Safety Index",
      "url": "https://futureoflife.org/ai-safety-index-summer-2025/",
      "resourceId": "df46edd6fa2078d1",
      "resourceTitle": "FLI AI Safety Index Summer 2025"
    },
    {
      "text": "Anthropic study",
      "url": "https://arxiv.org/abs/2401.05566",
      "resourceId": "e5c0904211c7d0cc",
      "resourceTitle": "Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training"
    },
    {
      "text": "Simple probes",
      "url": "https://www.anthropic.com/research/probes-catch-sleeper-agents",
      "resourceId": "72c1254d07071bf7",
      "resourceTitle": "Anthropic's follow-up research on defection probes"
    },
    {
      "text": "Greenblatt et al. 2024",
      "url": "https://www.anthropic.com/research/alignment-faking",
      "resourceId": "c2cfd72baafd64a9",
      "resourceTitle": "Anthropic's 2024 alignment faking study"
    },
    {
      "text": "Process supervision",
      "url": "https://arxiv.org/abs/2305.20050",
      "resourceId": "eea50d24e41938ed",
      "resourceTitle": "OpenAI's influential \"Let's Verify Step by Step\" study"
    },
    {
      "text": "AI Safety Index (Summer 2025)",
      "url": "https://futureoflife.org/ai-safety-index-summer-2025/",
      "resourceId": "df46edd6fa2078d1",
      "resourceTitle": "FLI AI Safety Index Summer 2025"
    },
    {
      "text": "February 2025 arXiv study",
      "url": "https://arxiv.org/html/2502.14870v1",
      "resourceId": "4e7f0e37bace9678",
      "resourceTitle": "Roman Yampolskiy"
    }
  ],
  "unconvertedLinkCount": 23,
  "convertedLinkCount": 26,
  "backlinkCount": 4,
  "hallucinationRisk": {
    "level": "medium",
    "score": 45,
    "factors": [
      "no-citations",
      "conceptual-content"
    ]
  },
  "entityType": "crux",
  "redundancy": {
    "maxSimilarity": 22,
    "similarPages": [
      {
        "id": "why-alignment-hard",
        "title": "Why Alignment Might Be Hard",
        "path": "/knowledge-base/debates/why-alignment-hard/",
        "similarity": 22
      },
      {
        "id": "agentic-ai",
        "title": "Agentic AI",
        "path": "/knowledge-base/capabilities/agentic-ai/",
        "similarity": 21
      },
      {
        "id": "language-models",
        "title": "Large Language Models",
        "path": "/knowledge-base/capabilities/language-models/",
        "similarity": 21
      },
      {
        "id": "solutions",
        "title": "AI Safety Solution Cruxes",
        "path": "/knowledge-base/cruxes/solutions/",
        "similarity": 21
      },
      {
        "id": "is-ai-xrisk-real",
        "title": "Is AI Existential Risk Real?",
        "path": "/knowledge-base/debates/is-ai-xrisk-real/",
        "similarity": 21
      }
    ]
  },
  "changeHistory": [
    {
      "date": "2026-03-15",
      "branch": "auto-update/2026-03-15",
      "title": "Auto-improve (standard): AI Accident Risk Cruxes",
      "summary": "Improved \"AI Accident Risk Cruxes\" via standard pipeline (496.7s). Quality score: 71. Issues resolved: Truncated content: The page ends mid-sentence in the Researc; EntityLink E1 ('adversarial-robustness') and E2 ('agentic-ai; EntityLink E1030 ('claude-3-5-sonnet') is used in Quick Asse.",
      "duration": "496.7s",
      "cost": "$5-8"
    },
    {
      "date": "2026-03-14",
      "branch": "auto-update/2026-03-14",
      "title": "Auto-improve (standard): AI Accident Risk Cruxes",
      "summary": "Improved \"AI Accident Risk Cruxes\" via standard pipeline (472.9s). Quality score: 71. Issues resolved: Truncated content: The page ends mid-sentence in the 'Resear; MDX Mermaid chart contains EntityLink JSX components inside ; Comparison pattern violation: description frontmatter contai.",
      "duration": "472.9s",
      "cost": "$5-8"
    },
    {
      "date": "2026-03-12",
      "branch": "auto-update/2026-03-12",
      "title": "Auto-improve (standard): AI Accident Risk Cruxes",
      "summary": "Improved \"AI Accident Risk Cruxes\" via standard pipeline (480.9s). Quality score: 74. Issues resolved: Frontmatter description field contains raw '<' character in ; Mermaid chart contains an EntityLink MDX component inside a ; Table in 'Sources and Resources > 2024-2025 Key Research' ha.",
      "duration": "480.9s",
      "cost": "$5-8"
    }
  ],
  "coverage": {
    "passing": 9,
    "total": 13,
    "targets": {
      "tables": 26,
      "diagrams": 3,
      "internalLinks": 51,
      "externalLinks": 32,
      "footnotes": 19,
      "references": 19
    },
    "actuals": {
      "tables": 26,
      "diagrams": 1,
      "internalLinks": 115,
      "externalLinks": 74,
      "footnotes": 0,
      "references": 42,
      "quotesWithQuotes": 0,
      "quotesTotal": 0,
      "accuracyChecked": 0,
      "accuracyTotal": 0
    },
    "items": {
      "summary": "green",
      "schedule": "green",
      "entity": "green",
      "editHistory": "green",
      "overview": "green",
      "tables": "green",
      "diagrams": "amber",
      "internalLinks": "green",
      "externalLinks": "green",
      "footnotes": "red",
      "references": "green",
      "quotes": "red",
      "accuracy": "red"
    },
    "editHistoryCount": 3,
    "ratingsString": "N:5.2 R:6.8 A:7.3 C:7.5"
  },
  "readerRank": 5,
  "researchRank": 1,
  "recommendedScore": 199.51
}

{
  "lesswrong": "https://www.lesswrong.com/tag/ai-risk"
}