AI systems do not exist in isolation. Organizations need to consider how the business processes and data flows built around AI systems can be designed in a way that reduces the business impact that a cybersecurity failure might cause. Where assurance on the security of underlying AI or on the effectiveness of defences is limited, it’s crucial to consider how any compromise might be overcome. This could include implementing additional controls outside the system itself, or reviewing what data should or should not be exposed to the AI. To enable such an end-to-end view, risks and controls need to be integrated into wider governance structures and enterprise risk management processes.Alongside shifting left and expanding right, any approach for mitigating the cybersecurity risks associated with AI adoption needs to consider how the technology will evolve and how business use will change over time. This should be facilitated via repeated re-evaluation of risks and controls, alongside frequent rehearsal and regular testing of the organization’s preparedness (e.g. war gaming, tabletop exercises, disaster recovery drills). This presents another opportunity to further integrate cyber risk assessment and intelligence capabilities into the resilience cycle and adjust testing strategies based on evolving AI risk profiles and threat actor developments observed across the industry. This means that leaders need to expand right, i.e. embed cyber resilience, and repeat. 2.4 Taking an enterprise view 2.3 Shift left, expand right and repeat The question of how to secure AI is closely related to a wider body of work related to AI safety. This work is a significant aspect of the AI Governance Alliance’s (AIGA’s) agenda. This approach promotes the need to “shift left”, i.e. implement safety guardrails early in the AI system life cycle (namely, at the building and pre-deployment stages) to mitigate related risks.16 As an example of safe and secure- by-design AI technologies, it mandates the use of processes that address inherent vulnerabilities in the AI systems and services being used and procured by organizations. Not all risks can be mitigated at the building and pre-deployment stages. It is not possible to eliminate all system vulnerabilities, and there will always be threat actors who will succeed in circumventing the mitigating measures in place. To complement the security-by-design practices that help organizations develop AI technologies securely and ethically, businesses need to implement cybersecurity practices that will protect AI systems once they are in use. This requires: –An understanding of the wider risks faced by businesses using and depending on AI –An understanding of the risks associated with the criticality of the data being processed –Effective operational cybersecurity capabilities to protect against these risks and detect attacks –Effective response and recovery processes to deal with incidents when they occur In short, organizations will need to both “shift left and expand right”.2.1 Shift left 2.2 Shift left and expand right Artificial Intelligence and Cybersecurity: Balancing Risks and Rewards 11

Artificial Intelligence and Cybersecurity Balancing Risks and Rewards 2025