Chapter 9 - Discussion Questions

9.1: What is the most difficult aspect of AI and cybersecurity GRC for you, and does it align with your sector? Discuss this with a group of peers across different sectors.

💡 It is important to discuss with other peers as their AI use cases can be different. Peers who operate across different jurisdictions can provide different insights.

9.2: Let us return to the discussion on how AI RMF is likely best used as a guide due to the differing governance requirements of different types of systems.

a. Let us re-examine the AI chatbot vs V2X system once more. Besides the fact that one of them is most likely an IT system, and another an amalgam of different types of systems (IT, IoT, OT), what architectural differences significantly impact the trade-offs required for their AI implementation?
💡 What is the worst that can happen to a hallucinating chatbot? How about a V2X system? Use the EU AI Act to classify the risk of these different AI systems.
b. Would there be a need to use different standards and norms to assess the V2X system as compared to the AI chatbot?
💡 A chatbot relies heavily on data privacy standards (like GDPR) and bias testing, while a V2X system demands rigorous safety-critical engineering standards (like ISO 26262) where human life is explicitly quantified in the risk model.
c. Based on your answer in (a), how do these trade-offs manifest in the decisions made on said AI systems?
💡 Because V2X cannot afford cloud latency or downtime, the AI models must be heavily compressed to run locally on the car's edge hardware. What is the impact on solution design?
d. Let us return to MLSecOps in Chapter 4. We mentioned how LLMs scale differently from traditional ML applications, necessitating a different SecOps approach. Use the AI RMF to explain how different AI systems affect risk management?
💡 Think about the use cases of ML applications and the range of outputs of ML applications. At that time, the topic on AI regulation has evolved rapidly since then. Why is that the case?

9.3: We mentioned the three lines model earlier in the chapter. The three lines pertain to the owners/managers of risk (first line), specialist support, advice, monitoring and challenge (second line), and internal audit (third line).

a. Use an example in the finance industry to explain how these may map to an AI systems management team.
💡 Who does, who checks and who audits?
b. Why is there a need for external audit beyond the three lines identified?
💡 Even a robust internal 3rd LoD reports to the same CEO who wants the product to succeed. This is a question that recaps the basics of auditing and the roles of the internal auditor with respect to an external auditor.
c. There have been attempts at incorporating the Three Lines model to cybersecurity, with mixed success. One such example was mooted by Prof. Jerry Perullo. In practice, how does this map to your organization?
💡 The Three Lines model in cybersecurity has evolved quite quickly, from their original purposes in audit. However, in present-day lean security teams, there are roles that often blur the lines, acting as both operators (1st LoD) and risk analysts (2nd LoD) simultaneously¹.
d. Would a similar LoD model for AI work?
💡 What AI properties are required for an auditor to successively audit an AI system? Review what makes an AI system fit for purpose, such as OECD's standards for AI systems. ²

¹ These roles are also called “1.5 LoD” roles. ² Organisation for Economic Co-operation and Development. (2024). OECD framework for the classification of AI systems. In OECD. https://www.oecd.org/en/publications/oecd-framework-for-the-classification-of-ai-systems_cb6d9eca-en.html

9.4: We can extend the topic on dual-use technology (DUT) to understanding the cybersecurity doctrines of different countries.

a. Which countries have explicitly advocated a purely defensive stance? How about those that have expressed the possibility of using cyber offense against adversaries?
💡 Research which countries currently have a “defend forward” doctrines as opposed to countries that strictly limit cyber operations to responding only after an attack occurs on their soil.
b. If you were a nation state, how would you view cybersecurity and AI topics from the perspective of national strategic interests?
💡 This is a rapidly developing topic. Keywords of interest include the AI supply chain (including microchips used for physical AI workloads), AI capabilities and sovereign AI, owing to supply chain risk.
c. Use your answer in (b) to comment on how the commercial world views cybersecurity and AI.
💡 Consider the statements that have been released by different commercial companies. Why did they say what they said, in an effort to lobby legislation or public opinion one way or the other?

9.5: We have seen how the EU AI Act, ISO and CoSAI answer different aspects of the governance framework for AI systems.

a. To what extent do you consider them sufficient?
💡 Read the frameworks. Then view them from the strategic, tactical and operational lens.
b. Other countries take reference from the EU AI Act when drafting their own countries’ legislation. What examples have you seen, and to what extent do they affect how AI systems in these countries need to be designed to be compliant in said jurisdictions?
💡 To what extent does the “Brussels Effect” affect the AI space³?

³ Bradford, A. (2012). The brussels effect. Nw. UL Rev., 107, 1.