AI Compounding Risks Analysis Model

compounding-risks-analysis (E63)

← Back to pagePath: /knowledge-base/models/compounding-risks-analysis/

Page Metadata

{
  "id": "compounding-risks-analysis",
  "numericId": null,
  "path": "/knowledge-base/models/compounding-risks-analysis/",
  "filePath": "knowledge-base/models/compounding-risks-analysis.mdx",
  "title": "Compounding Risks Analysis",
  "quality": 60,
  "importance": 67,
  "contentFormat": "article",
  "tractability": null,
  "neglectedness": null,
  "uncertainty": null,
  "causalLevel": null,
  "lastUpdated": "2025-12-26",
  "llmSummary": "Mathematical framework quantifying how AI risks compound beyond additive effects through four mechanisms (multiplicative probability, severity multiplication, defense negation, nonlinear effects), with racing+deceptive alignment showing 3-8% catastrophic probability and interaction coefficients of 2-10x. Provides specific cost-effectiveness estimates for interventions targeting compound pathways ($1-4M per 1% risk reduction) and demonstrates systematic 2-5x underestimation by traditional additive models.",
  "structuredSummary": null,
  "description": "Mathematical framework showing how AI risks compound beyond additive effects through four mechanisms (multiplicative probability, severity multiplication, defense negation, nonlinear effects). Racing+deceptive alignment combinations show 3-8% catastrophic probability, with interaction coefficients of 2-10x requiring systematic intervention targeting compound pathways.",
  "ratings": {
    "focus": 8.5,
    "novelty": 5,
    "rigor": 4.5,
    "completeness": 7,
    "concreteness": 7.5,
    "actionability": 6
  },
  "category": "models",
  "subcategory": "analysis-models",
  "clusters": [
    "ai-safety",
    "governance"
  ],
  "metrics": {
    "wordCount": 1792,
    "tableCount": 16,
    "diagramCount": 2,
    "internalLinks": 54,
    "externalLinks": 0,
    "footnoteCount": 0,
    "bulletRatio": 0.08,
    "sectionCount": 30,
    "hasOverview": true,
    "structuralScore": 12
  },
  "suggestedQuality": 80,
  "updateFrequency": 90,
  "evergreen": true,
  "wordCount": 1792,
  "unconvertedLinks": [],
  "unconvertedLinkCount": 0,
  "convertedLinkCount": 28,
  "backlinkCount": 3,
  "redundancy": {
    "maxSimilarity": 15,
    "similarPages": [
      {
        "id": "risk-interaction-matrix",
        "title": "Risk Interaction Matrix Model",
        "path": "/knowledge-base/models/risk-interaction-matrix/",
        "similarity": 15
      },
      {
        "id": "risk-interaction-network",
        "title": "Risk Interaction Network",
        "path": "/knowledge-base/models/risk-interaction-network/",
        "similarity": 15
      },
      {
        "id": "ai-risk-portfolio-analysis",
        "title": "AI Risk Portfolio Analysis",
        "path": "/knowledge-base/models/ai-risk-portfolio-analysis/",
        "similarity": 14
      },
      {
        "id": "capability-alignment-race",
        "title": "Capability-Alignment Race Model",
        "path": "/knowledge-base/models/capability-alignment-race/",
        "similarity": 14
      },
      {
        "id": "corrigibility-failure-pathways",
        "title": "Corrigibility Failure Pathways",
        "path": "/knowledge-base/models/corrigibility-failure-pathways/",
        "similarity": 14
      }
    ]
  }
}

Entity Data

{
  "id": "compounding-risks-analysis",
  "type": "model",
  "title": "AI Compounding Risks Analysis Model",
  "description": "This model analyzes how risks compound beyond additive effects. Key combinations include racing+concentration (40-60% coverage needed) and mesa-optimization+scheming (2-6% catastrophic probability).",
  "tags": [
    "risk-interactions",
    "compounding-effects",
    "systems-thinking"
  ],
  "relatedEntries": [
    {
      "id": "risk-interaction-matrix",
      "type": "model",
      "relationship": "related"
    },
    {
      "id": "risk-cascade-pathways",
      "type": "model",
      "relationship": "related"
    }
  ],
  "sources": [],
  "lastUpdated": "2025-12",
  "customFields": [
    {
      "label": "Model Type",
      "value": "Systems Analysis"
    },
    {
      "label": "Scope",
      "value": "Multi-Risk Interactions"
    },
    {
      "label": "Key Insight",
      "value": "Combined risks often exceed the sum of individual risks due to non-linear interactions"
    }
  ]
}

Canonical Facts (0)

No facts for this entity

External Links

No external links

Backlinks (3)

id	title	type	relationship
ai-risk-portfolio-analysis	AI Risk Portfolio Analysis	model	related
risk-cascade-pathways	AI Risk Cascade Pathways Model	model	related
risk-interaction-network	AI Risk Interaction Network Model	model	related

Frontmatter

{
  "title": "Compounding Risks Analysis",
  "description": "Mathematical framework showing how AI risks compound beyond additive effects through four mechanisms (multiplicative probability, severity multiplication, defense negation, nonlinear effects). Racing+deceptive alignment combinations show 3-8% catastrophic probability, with interaction coefficients of 2-10x requiring systematic intervention targeting compound pathways.",
  "sidebar": {
    "order": 51
  },
  "quality": 60,
  "lastEdited": "2025-12-26",
  "ratings": {
    "focus": 8.5,
    "novelty": 5,
    "rigor": 4.5,
    "completeness": 7,
    "concreteness": 7.5,
    "actionability": 6
  },
  "importance": 67.5,
  "update_frequency": 90,
  "llmSummary": "Mathematical framework quantifying how AI risks compound beyond additive effects through four mechanisms (multiplicative probability, severity multiplication, defense negation, nonlinear effects), with racing+deceptive alignment showing 3-8% catastrophic probability and interaction coefficients of 2-10x. Provides specific cost-effectiveness estimates for interventions targeting compound pathways ($1-4M per 1% risk reduction) and demonstrates systematic 2-5x underestimation by traditional additive models.",
  "todos": [
    "Complete 'Conceptual Framework' section",
    "Complete 'Quantitative Analysis' section (8 placeholders)",
    "Complete 'Strategic Importance' section"
  ],
  "clusters": [
    "ai-safety",
    "governance"
  ],
  "subcategory": "analysis-models",
  "entityType": "model"
}

Raw MDX Source

---
title: Compounding Risks Analysis
description: Mathematical framework showing how AI risks compound beyond additive effects through four mechanisms (multiplicative probability, severity multiplication, defense negation, nonlinear effects). Racing+deceptive alignment combinations show 3-8% catastrophic probability, with interaction coefficients of 2-10x requiring systematic intervention targeting compound pathways.
sidebar:
  order: 51
quality: 60
lastEdited: "2025-12-26"
ratings:
  focus: 8.5
  novelty: 5
  rigor: 4.5
  completeness: 7
  concreteness: 7.5
  actionability: 6
importance: 67.5
update_frequency: 90
llmSummary: Mathematical framework quantifying how AI risks compound beyond additive effects through four mechanisms (multiplicative probability, severity multiplication, defense negation, nonlinear effects), with racing+deceptive alignment showing 3-8% catastrophic probability and interaction coefficients of 2-10x. Provides specific cost-effectiveness estimates for interventions targeting compound pathways ($1-4M per 1% risk reduction) and demonstrates systematic 2-5x underestimation by traditional additive models.
todos:
  - Complete 'Conceptual Framework' section
  - Complete 'Quantitative Analysis' section (8 placeholders)
  - Complete 'Strategic Importance' section
clusters:
  - ai-safety
  - governance
subcategory: analysis-models
entityType: model
---
import {DataInfoBox, KeyQuestions, Mermaid, R, EntityLink} from '@components/wiki';

<DataInfoBox entityId="E63" ratings={frontmatter.ratings} />

## Overview

When multiple AI risks occur simultaneously, their combined impact often dramatically exceeds simple addition. This mathematical framework analyzes how <EntityLink id="E239">racing dynamics</EntityLink>, <EntityLink id="E93">deceptive alignment</EntityLink>, and <EntityLink id="E179">lock-in scenarios</EntityLink> interact through four compounding mechanisms. The central insight: a world with three moderate risks isn't 3x as dangerous as one with a single risk—it can be 10-20x more dangerous due to multiplicative interactions.

Analysis of high-risk combinations reveals that racing+deceptive alignment scenarios carry 3-8% catastrophic probability, while <EntityLink id="E197">mesa-optimization</EntityLink>+<EntityLink id="E274">scheming</EntityLink> pathways show 2-6% existential risk. Traditional additive risk models systematically underestimate total danger by factors of 2-5x because they ignore how risks amplify each other's likelihood, severity, and defensive evasion.

The framework provides quantitative interaction coefficients (α values of 2-10x for severity multiplication, 3-6x for probability amplification) and mathematical models to correct this systematic underestimation. This matters for resource allocation: reducing compound pathways often provides higher leverage than addressing individual risks in isolation.

## Risk Compounding Assessment

| Risk Combination | Interaction Type | Compound Probability | Severity Multiplier | Confidence Level |
|------------------|------------------|---------------------|-------------------|------------------|
| Racing + Deceptive Alignment | Probability multiplication | 15.8% vs 4.5% baseline | 3.5x | Medium |
| Deceptive + <EntityLink id="E189">Lock-in</EntityLink> | Severity multiplication | 8% | 8-10x | Medium |
| <EntityLink id="E133">Expertise Atrophy</EntityLink> + Corrigibility Failure | Defense negation | Variable | 3.3x | Medium-High |
| Mesa-opt + Scheming | Nonlinear combined | 2-6% catastrophic | Discontinuous | Medium |
| <EntityLink id="E119">Epistemic Collapse</EntityLink> + Democratic Failure | Threshold crossing | 8-20% | Qualitative change | Low |

## Compounding Mechanisms Framework

### Mathematical Foundation

Traditional additive models dramatically underestimate compound risk:

| Model Type | Formula | Typical Underestimate | Use Case |
|------------|---------|----------------------|----------|
| **Naive Additive** | $R_{total} = R_1 + R_2 + ... + R_n$ | 2-5x underestimate | Individual risk planning |
| **Multiplicative** | $R_{total} = 1 - \prod_i(1 - R_i) \times IF$ | 1.5-3x underestimate | Overlapping vulnerabilities |
| **Synergistic (Recommended)** | $R_{total} = \sum_i R_i + \sum_{i<j} \alpha_{ij} R_i R_j + ...$ | Baseline accuracy | Compound risk assessment |

**Synergistic Model (Full Specification)**:
$$\text{Total Risk} = \sum_{i} R_i + \sum_{i<j} \alpha_{ij} R_i R_j + \sum_{i<j<k} \beta_{ijk} R_i R_j R_k$$

Where α coefficients represent pairwise interaction strength and β coefficients capture three-way interactions.

### Type 1: Multiplicative Probability

When Risk A increases the likelihood of Risk B:

| Scenario | P(Mesa-opt) | P(Deceptive \| Mesa-opt) | Combined Probability | Compounding Factor |
|----------|-------------|--------------------------|---------------------|-------------------|
| Baseline (no racing) | 15% | 30% | 4.5% | 1x |
| Moderate racing | 25% | 40% | 10% | 2.2x |
| Intense racing | 35% | 45% | 15.8% | 3.5x |
| Extreme racing | 50% | 55% | 27.5% | 6.1x |

**Mechanism**: <EntityLink id="E239">Racing dynamics</EntityLink> compress safety timelines → inadequate testing → higher probability of <EntityLink id="E197">mesa-optimization</EntityLink> → higher probability of <EntityLink id="E93">deceptive alignment</EntityLink>.

### Type 2: Severity Multiplication

When combined risks produce worse outcomes than the sum of individual impacts:

| Individual Risk | Standalone Impact | Combined with Lock-in | Severity Multiplier |
|-----------------|-------------------|----------------------|-------------------|
| Deceptive Alignment | High (misaligned optimization) | Catastrophic (trapped forever) | 8-10x |
| <EntityLink id="E80">Corrigibility Failure</EntityLink> | Medium (hard to correct) | Severe (permanent embedding) | 4-6x |
| <EntityLink id="E295">Sycophancy</EntityLink> | Low (biased advice) | High (institutionalized bias) | 3-5x |

### Type 3: Defense Negation

When Risk A disables safeguards against Risk B:

| <EntityLink id="E159">Human Expertise</EntityLink> Level | P(Detect Deceptive Behavior) | P(Catastrophe \| Undetected) | Total Catastrophe Risk |
|-----------------------|-------------------------------|------------------------------|----------------------|
| Full expertise maintained | 60% | 15% | 6% |
| Moderate <EntityLink id="E133">expertise atrophy</EntityLink> | 35% | 30% | 19.5% (3.3x) |
| Severe expertise atrophy | 15% | 50% | 42.5% (7x) |

### Type 4: Nonlinear Combined Effects

When interactions produce qualitatively different outcomes:

| Combined Stressors | Individual Effect | Compound Effect | Threshold Behavior |
|-------------------|------------------|-----------------|-------------------|
| Epistemic degradation alone | Manageable stress on institutions | - | Linear response |
| Political polarization alone | Manageable stress on institutions | - | Linear response |
| **Both together** | - | Democratic system failure | Phase transition |

<Mermaid chart={`flowchart TD
    A[Individual Risks] --> B[Additive Model<br/>R₁ + R₂ + R₃]
    A --> C[Compound Model<br/>Σ + ΣΣα + ΣΣΣβ]
    
    B --> D[Underestimate<br/>2-5x too low]
    C --> E[Accurate Assessment<br/>Captures interactions]
    
    F[Racing Dynamics] --> G[Higher Mesa-opt Probability]
    G --> H[Higher Deceptive Alignment]
    H --> I[Lock-in Risk]
    I --> J[Catastrophic Outcome<br/>3-8% probability]
    
    style D fill:#ffcccc
    style E fill:#ccffcc
    style J fill:#ff9999
`} />

## High-Risk Compound Combinations

### Critical Interaction Matrix

| Risk A | Risk B | Interaction Strength (α) | Combined Catastrophe Risk | Evidence Source |
|--------|--------|-------------------------|--------------------------|-----------------|
| Racing + Deceptive Alignment | 3.0-5.0 | 3-8% | <R id="cd3035dbef6c7b5b">Amodei et al. (2016)</R> |
| Deceptive + Lock-in | 5.0-10.0 | 8-15% | <R id="64ad308db00b3ce7">Carlsmith (2021)</R> |
| <EntityLink id="E197">Mesa-optimization</EntityLink> + <EntityLink id="E274">Scheming</EntityLink> | 3.0-6.0 | 2-6% | <R id="c4858d4ef280d8e6">Hubinger et al. (2019)</R> |
| Expertise Atrophy + Corrigibility Failure | 2.0-4.0 | 5-12% | <R id="eecd7c0e9ebb9cbe">RAND Corporation</R> |
| <EntityLink id="E374">Concentration</EntityLink> + <EntityLink id="E30">Authoritarian Tools</EntityLink> | 3.0-5.0 | 5-12% | <R id="a306e0b63bdedbd5">Center for AI Safety</R> |

### Three-Way Compound Scenarios

| Scenario | Risk Combination | Compound Probability | Recovery Likelihood | Assessment |
|----------|------------------|---------------------|-------------------|------------|
| **Technical Cascade** | Racing + Mesa-opt + Deceptive | 3-8% | Very Low | Most dangerous technical pathway |
| **Structural Lock-in** | Deceptive + Lock-in + Authoritarian | 5-12% | Near-zero | Permanent misaligned control |
| **Oversight Failure** | Sycophancy + Expertise + Corrigibility | 5-15% | Low | No human check on behavior |
| **Coordination Collapse** | Epistemic + Trust + Democratic | 8-20% | Medium | Civilization coordination failure |

## Quantitative Risk Calculation

### Worked Example: Racing + Deceptive + Lock-in

**Base Probabilities**:
- Racing dynamics (R₁): 30%
- Deceptive alignment (R₂): 15% 
- Lock-in scenario (R₃): 20%

**Interaction Coefficients**:
- α₁₂ = 2.0 (racing increases deceptive probability)
- α₁₃ = 1.5 (racing increases lock-in probability)  
- α₂₃ = 3.0 (deceptive alignment strongly increases lock-in severity)

**Calculation**:
$$\text{P(Compound)} = R_1 + R_2 + R_3 + \alpha_{12}R_1R_2 + \alpha_{13}R_1R_3 + \alpha_{23}R_2R_3$$

$$= 0.30 + 0.15 + 0.20 + 2.0(0.045) + 1.5(0.06) + 3.0(0.03)$$

$$= 0.65 + 0.09 + 0.09 + 0.09 = 0.92$$

**Interpretation**: 92% probability that at least one major compound effect occurs, with severity multiplication making outcomes far worse than individual risks would suggest.

### Scenario Probability Analysis

| Scenario | 2030 Probability | 2040 Probability | Compound Risk Level | Primary Drivers |
|----------|------------------|------------------|-------------------|-----------------|
| **Correlated Realization** | 8% | 15% | Critical (0.9+) | Competitive pressure drives all risks |
| **Gradual Compounding** | 25% | 40% | High (0.6-0.8) | Slow interaction buildup |
| **Successful Decoupling** | 15% | 25% | Moderate (0.3-0.5) | Interventions break key links |
| **Threshold Cascade** | 12% | 20% | Variable | Sudden phase transition |

**Expected Compound Risk by 2040**: 
$$E[Risk] = 0.15(0.9) + 0.40(0.7) + 0.25(0.4) + 0.20(0.65) = 0.645$$

## Current State & Trajectory

### Present Compound Risk Indicators

| Indicator | Current Level | Trend | 2030 Projection | Key Evidence |
|-----------|---------------|-------|-----------------|--------------|
| Racing intensity | Moderate-High | ↗ Increasing | High | <R id="5fa46de681ff9902">AI lab competition</R>, <R id="120adc539e2fa558">compute scaling</R> |
| Technical risk correlation | Medium | ↗ Increasing | Medium-High | <R id="2e0c662574087c2a">Mesa-optimization research</R> |
| Lock-in pressure | Low-Medium | ↗ Increasing | Medium-High | <R id="456759f23f47ea0a">Market concentration</R> |
| Expertise preservation | Medium | ↘ Decreasing | Low-Medium | <R id="0a17f30e99091ebf">RAND workforce analysis</R> |
| Defensive capabilities | Medium | → Stable | Medium | <R id="3b9fda03b8be71dc">AI safety funding</R> |

### Key Trajectory Drivers

**Accelerating Factors**:
- Geopolitical competition intensifying AI race
- <EntityLink id="E272">Scaling laws</EntityLink> driving capability advances
- Economic incentives favoring rapid deployment
- Regulatory lag behind capability development

**Mitigating Factors**:
- Growing AI safety community and funding
- Industry <EntityLink id="E369">voluntary commitments</EntityLink>
- International coordination efforts (<EntityLink id="E279">Seoul Declaration</EntityLink>)
- Technical progress on <EntityLink id="E176">interpretability</EntityLink> and alignment

## High-Leverage Interventions

### Intervention Effectiveness Matrix

| Intervention | Compound Pathways Addressed | Risk Reduction | Annual Cost | Cost-Effectiveness |
|--------------|----------------------------|----------------|-------------|-------------------|
| **Reduce racing dynamics** | Racing × all technical risks | 40-60% | \$500M-1B | \$2-4M per 1% reduction |
| **Preserve human expertise** | Expertise × all oversight risks | 30-50% | \$200M-500M | \$1-3M per 1% reduction |
| **Prevent lock-in** | Lock-in × all structural risks | 50-70% | \$300M-600M | \$1-2M per 1% reduction |
| **Maintain epistemic health** | Epistemic × democratic risks | 30-50% | \$100M-300M | \$1-2M per 1% reduction |
| **International coordination** | Racing × concentration × authoritarian | 30-50% | \$200M-500M | \$1-3M per 1% reduction |

### Breaking Compound Cascades

<Mermaid chart={`flowchart TD
    A[Racing Dynamics] -->|α=2.0| B[Technical Risks]
    B -->|α=4.0| C[Lock-in Effects]
    C -->|α=3.5| D[Structural Risks]
    
    I1[Slow racing] -.->|Intervention 1| A
    I2[Preserve expertise] -.->|Intervention 2| B
    I3[Prevent lock-in] -.->|Intervention 3| C
    I4[Democratic safeguards] -.->|Intervention 4| D
    
    style A fill:#ffcccc
    style B fill:#ffcccc  
    style C fill:#ffcccc
    style D fill:#ff9999
    style I1 fill:#ccffcc
    style I2 fill:#ccffcc
    style I3 fill:#ccffcc
    style I4 fill:#ccffcc
`} />

**Strategic Insights**:
- Early intervention (before racing intensifies) provides highest leverage
- Breaking any major pathway (racing→technical, technical→lock-in) dramatically reduces compound risk
- Preserving human oversight capabilities acts as universal circuit breaker

## Key Uncertainties & Cruxes

### Critical Unknowns

<KeyQuestions
  questions={[
    "Are interaction coefficients stable across different AI capability levels?",
    "Which three-way combinations pose the highest existential risk?", 
    "Can we detect threshold approaches before irreversible cascades begin?",
    "Do positive interactions (risks that reduce each other) meaningfully offset negative ones?",
    "How do defensive interventions interact - do they compound positively?"
  ]}
/>

### Expert Disagreement Areas

| Uncertainty | Optimistic View | Pessimistic View | Current Evidence |
|-------------|-----------------|------------------|------------------|
| **Interaction stability** | Coefficients decrease as AI improves | Coefficients increase with capability | Mixed signals from capability research |
| **Threshold existence** | Gradual degradation, no sharp cutoffs | Clear tipping points exist | Limited historical analogies |
| **Intervention effectiveness** | Targeted interventions highly effective | System too complex for reliable intervention | Early positive results from <EntityLink id="E252">responsible scaling</EntityLink> |
| **Timeline urgency** | Compound effects emerge slowly (10+ years) | Critical combinations possible by 2030 | <EntityLink id="E4">AGI timeline uncertainty</EntityLink> |

## Limitations & Model Validity

### Methodological Constraints

**Interaction coefficient uncertainty**: α values are based primarily on expert judgment and theoretical reasoning rather than empirical measurement. Different analysts could reasonably propose coefficients differing by 2-3x, dramatically changing risk estimates. The <R id="a306e0b63bdedbd5">Center for AI Safety</R> and <R id="1593095c92d34ed8">Future of Humanity Institute</R> have noted similar calibration challenges in compound risk assessment.

**Higher-order effects**: The model focuses on pairwise interactions but real catastrophic scenarios likely require 4+ simultaneous risks. The <EntityLink id="E12" /> suggests higher-order terms may dominate in extreme scenarios.

**Temporal dynamics**: Risk probabilities and interaction strengths evolve as AI capabilities advance. Racing dynamics mild today may intensify rapidly; interaction effects manageable at current capability levels may become overwhelming as systems become more powerful.

### Validation Challenges

| Challenge | Impact | Mitigation Strategy |
|-----------|--------|-------------------|
| **Pre-catastrophe validation impossible** | Cannot test model accuracy without experiencing failures | Use historical analogies, stress-test assumptions |
| **Expert disagreement on coefficients** | 2-3x uncertainty in final estimates | Report ranges, sensitivity analysis |
| **Intervention interaction effects** | Reducing one risk might increase others | Model defensive interactions explicitly |
| **Threshold precision claims** | False precision in "tipping point" language | Emphasize continuous degradation |

## Sources & Resources

### Academic Literature

| Source | Focus | Key Finding | Relevance |
|--------|-------|-------------|-----------|
| <R id="cd3035dbef6c7b5b">Amodei et al. (2016)</R> | AI safety problems | Risk interactions in reward systems | High - foundational framework |
| <R id="64ad308db00b3ce7">Carlsmith (2021)</R> | Power-seeking AI | Lock-in mechanism analysis | High - severity multiplication |
| <R id="c4858d4ef280d8e6">Hubinger et al. (2019)</R> | Mesa-optimization | Deceptive alignment pathways | High - compound technical risks |
| <R id="28240d2bdf0f01d5">Russell (2019)</R> | AI alignment | Compound failure modes | Medium - conceptual framework |

### Research Organizations

| Organization | Contribution | Key Publications |
|--------------|-------------|------------------|
| <R id="afe2508ac4caf5ee">Anthropic</R> | Compound risk research | <R id="683aef834ac1612a">Constitutional AI</R> |
| <R id="a306e0b63bdedbd5">Center for AI Safety</R> | Risk interaction analysis | <R id="470ac236ca26008c">AI Risk Statement</R> |
| <R id="0a17f30e99091ebf">RAND Corporation</R> | Expertise atrophy studies | <R id="eecd7c0e9ebb9cbe">AI Workforce Analysis</R> |
| <R id="1593095c92d34ed8">Future of Humanity Institute</R> | Existential risk modeling | <R id="902320774d220a6c">Global Catastrophic Risks</R> |

### Policy & Governance

| Resource | Focus | Application |
|----------|-------|-------------|
| <R id="54dbc15413425997">NIST AI Risk Management Framework</R> | Risk assessment methodology | Compound risk evaluation |
| <R id="fdf68a8f30f57dee">UK AI Safety Institute</R> | Safety evaluation | Interaction testing protocols |
| <R id="3afc13e40d7102b1">EU AI Act</R> | Regulatory framework | Compound risk regulation |