CORTEX: A New Framework for Quantifying AI System Risks

TLDR: CORTEX is a novel, multi-layered risk scoring framework designed to assess and quantify vulnerabilities in operational AI systems. Developed from an analysis of over 1,200 AI incidents, it categorizes failures into 29 technical groups and uses a five-tier architecture. This architecture combines utility-adjusted likelihood and impact, governance and contextual overlays, technical exposure scores, environmental and residual modifiers, and probabilistic modeling via Bayesian aggregation and Monte Carlo simulation. CORTEX aims to bridge the gap between qualitative AI governance principles and quantitative operational risk management, providing a dynamic and uncertainty-aware approach to identify, prioritize, and mitigate AI risks.

As Artificial Intelligence (AI) systems become increasingly integrated into critical sectors like healthcare, finance, education, and infrastructure, the potential for system failures and their impact has grown significantly. These failures are no longer theoretical but represent a recurring, systemic risk. To address this evolving challenge, a new framework called CORTEX (Composite Overlay for Risk Tiering and Exposure) has been introduced.

The Challenge of AI Risk Management

Current AI governance frameworks, such as the EU AI Act, NIST AI Risk Management Framework (AI RMF), and ISO/IEC 42001, provide high-level guidance on responsible AI. However, they often fall short in offering concrete mechanisms for operationalizing risk scoring and benchmarking the severity of AI system failures. Unlike cybersecurity, which benefits from quantitative tools like CVSS and FAIR, the AI domain has lacked a comprehensive system to map real-world incidents to technical vulnerabilities, lifecycle traceability, and governance preparedness.

Existing approaches have several limitations:

• **Absence of a Unified Risk Taxonomy and Scoring:** There hasn’t been a standardized way to categorize technical AI vulnerabilities and assign reproducible risk scores using empirical and contextual data.
• **Audit Misalignment:** Without quantitative outputs, it’s difficult to integrate AI risk assessments into formal processes like conformity assessments or internal audits.
• **No Volatility Analysis:** Most risk models treat AI risk as static, ignoring the dynamic and non-deterministic nature of AI systems, which can exhibit performance volatility, tail-risk escalation, and score drift over time.

Introducing CORTEX: A Multi-Layered Approach to AI Risk Scoring

CORTEX is a multi-layered risk scoring framework designed to assess and score AI system vulnerabilities. It was developed based on an empirical analysis of over 1,200 incidents documented in the AI Incident Database (AIID), categorizing failure modes into 29 technical vulnerability groups. The framework’s core strength lies in its five-tier architecture, which combines various factors to produce a comprehensive risk score:

1. **Utility-Adjusted Likelihood × Impact Calculations:** This layer assesses the probability and severity of an incident, with a non-linear adjustment to reflect a greater sensitivity to high-severity events.
2. **Governance + Contextual Overlays:** These modifiers align with regulatory frameworks like the EU AI Act, NIST RMF, and OECD principles, adjusting the risk score based on the specific context and regulatory environment of the AI system.
3. **Technical Surface Scores:** This covers exposure vectors such as model drift, traceability gaps, and adversarial risks, evaluating the inherent instability or exposure of a vulnerability.
4. **Environmental and Residual Modifiers:** These factors account for deployment context (e.g., public accessibility, use in critical domains) and the degree of unmitigated risk remaining after safeguards are in place.
5. **Final Layered Assessment via Bayesian Risk Aggregation and Monte Carlo Simulation:** This crucial layer models volatility and long-tail risks by treating each parameter as a probability distribution, providing a dynamic and uncertainty-aware risk assessment.

How CORTEX Works in Practice

The CORTEX framework begins by classifying AI system failures into a multi-tiered taxonomy, derived from real-world incidents. This taxonomy includes 7 high-level domains and 29 grouped vulnerabilities, further broken down into over 120 distinct failure types.

Each vulnerability is then scored through the five layers:

• **Baseline Risk:** An initial score is calculated using utility-transformed Likelihood (based on incident frequency) and Impact (calibrated for sectoral harm).
• **Contextual and Governance Modifiers:** Scores are adjusted based on the sensitivity of the deployment domain (e.g., systems for children or medical purposes receive higher ‘Contextual’ scores) and the level of regulatory oversight (e.g., high-risk systems under the EU AI Act receive higher ‘Governance’ scores).
• **Technical Surface Scoring:** This modifier accounts for technical exposure, such as bias severity, explainability gaps, model drift, and security vulnerabilities.
• **Environmental and Residual Risk Modifiers:** These consider factors like public accessibility and the effectiveness of existing mitigations. Systems with poor logging or human-in-the-loop oversight will have higher ‘Residual Risk’.
• **Probabilistic Risk Modeling:** This advanced layer uses Bayesian aggregation and Monte Carlo simulations to model uncertainty. Instead of static values, parameters are treated as probability distributions, allowing for the generation of median (P50) and upper-bound (P90) risk scores, along with standard deviation to measure volatility. This helps identify risks that might escalate under uncertain conditions.

The final CORTEX score is a weighted composite of these layers, normalized to a range, allowing for clear interpretation into risk tiers: Critical (≥ 0.85), High (0.70–0.84), Moderate (0.50–0.69), Low (0.30–0.49), and Minimal (< 0.30).

Also Read:

• Corporate AI Risk Reporting: A Look Inside SEC Filings
• Understanding Cyber Threats: An AI Approach to Vulnerability Mapping

Real-World Application and Insights

The paper demonstrates CORTEX’s application to various AI vulnerabilities. For instance, ‘Discriminatory Outcomes’, ‘Prompt Injection & Manipulation’, and ‘PII Leakage’ consistently rank as high-risk, even with probabilistic simulations. These vulnerabilities demand immediate attention and robust mitigation strategies. On the other hand, ‘UI-Induced Overtrust’ and ‘Feedback Loop Abuse’ might appear as lower risk in general scenarios but show sensitivity to contextual and governance shifts, highlighting their potential to escalate in specific, sensitive deployments.

CORTEX provides a dynamic and simulation-informed approach to AI risk management, moving beyond static assessments. It helps organizations prioritize mitigation efforts, align with compliance thresholds, and allocate resources effectively. By integrating technical, contextual, and probabilistic dimensions, CORTEX offers a transparent and traceable foundation for security reviews, model validations, regulatory assessments, and operational audits.

This framework represents a significant step towards making AI risk governance quantifiable, traceable, and adaptable, ensuring that AI systems are not only innovative but also responsible and trustworthy. For more in-depth information, you can read the full research paper: CORTEX: Composite Overlay for Risk Tiering and Exposure in Operational AI Systems.