The FCA published its multi-firm review on frontier AI and cyber resilience, while the BoE published practical considerations on frontier AI for cyber defence, on 2 September 2026.
Drawing on firms’ experience and industry discussions, the publications provide practical insights into the opportunities and cyber risks associated with frontier AI. They consider the surrounding governance, controls and operating environments, and how firms can manage a potentially greater volume of vulnerability findings.
Financial Stability Board (FSB) Chair Andrew Bailey also highlighted frontier AI’s potential to accelerate cyber risk and amplify disruption in a letter to G20 Finance Ministers and Central Bank Governors on 28 August 2026.
In this article
What does this mean?
Frontier AI is becoming a test of organisational resilience
The FCA’s review found that frontier AI can accelerate the discovery, validation and prioritisation of cyber vulnerabilities. However, faster discovery may expose weaknesses in firms’ governance, engineering resources, change processes and remediation capacity. The principal constraint may therefore be less the model’s capability than firms’ ability to validate, prioritise and act on its outputs safely.
Even after expert validation and triage, genuine vulnerabilities could put pressure on remediation teams and change-management processes. Bottlenecks may arise across patch testing, emergency change and evidence of closure, while firms must continue delivering important business services. The FCA notes that governance forums and senior leaders may need greater visibility of the impact on vulnerability registers, supplier dependencies and operational resilience.
Effective cyber defence depends on the design of the harness around the model
Drawing on discussions through the Frontier AI Information Sharing Forum (FAISF), the BoE shares practical considerations for the harness surrounding a frontier AI model. This includes the tools, workflows, controls, data and operating environment through which the model is used. Harness design shapes what the model can access and do, how its outputs are checked, and how findings reach cyber and engineering teams. A well-designed harness can make those findings more usable, explainable and actionable.
The BoE identifies a range of potential approaches, including internally developed tools, vendor platforms, managed services and open-source components. Each presents different trade-offs relating to flexibility, transparency, integration, maintenance and supportability. The BoE notes that modular or orchestration layers could help firms coordinate specialist tools, compare outputs and reduce early dependence on a single model or supplier. Scaling frontier AI for cyber defence may also depend on operating models that bring together red teaming, software engineering, AI literacy, security architecture and operational cyber expertise.
Controls and vulnerability management may need to evolve
The FCA emphasises that human judgement remains critical, while the BoE also identifies human review and oversight as important elements of harness design. Firms continue to rely on specialist expertise to validate findings, assess relevance, set priorities and make risk-based decisions. Governance can be embedded through measures including use-case restrictions, controlled access, network isolation, segregated environments, monitoring and approval workflows.
The BoE also highlights the trade-off between providing models with sufficient organisational context and protecting sensitive systems, data and intellectual property. Firms are exploring isolated, sandboxed and production-like environments to manage this risk.
The FCA notes that frontier AI may combine lower-rated weaknesses to create alternative attack paths. Prioritisation may therefore need to look beyond individual severity ratings and consider exploitability, exposure, compensating controls, system dependencies and business-service impact.
Third-party dependencies create a system-wide dimension
The FCA highlights the growing importance of supplier preparedness, cloud dependencies, software supply chains and shared infrastructure. It encourages firms to consider how suppliers use AI-enabled vulnerability discovery, validate and communicate findings, and manage increased remediation demands.
The FCA also sets out a series of questions to help firms assess their readiness across ownership and escalation, specialist review, remediation capacity, risk-based prioritisation, asset and supplier visibility, and the resilience of important business services.
Andrew Bailey, FSB Chair, also warned in his letter to G20 Finance Ministers and Central Bank Governors that frontier-AI-enabled cyber disruption could spread across jurisdictions through common technology providers, shared infrastructure and cross-border financial activity. The letter calls for stronger vulnerability-management, response and recovery capabilities, including preparation for simultaneous disruption and 'bare metal' recovery, where critical systems and data must be rebuilt from a clean foundation following a severe cyber incident.
A Financial Stability Institute paper, published by the Bank for International Settlements on 9 September 2026, similarly identifies international convergence around reinforcing existing cyber and operational resilience frameworks, with greater emphasis on faster decision-making, patching, response and recovery.
What do firms need to do?
Assess the full frontier AI harness, not only the underlying model.
Stress-test vulnerability management and remediation capacity.
Strengthen governance, human oversight and third-party resilience.
While the publications do not introduce new regulatory requirements, they signal growing regulatory focus on frontier AI’s implications for cyber and operational resilience and provide practical insights.
Firms should assess the tools, workflows, data, permissions and operating environments surrounding frontier AI models, with clear ownership, proportionate guardrails, controlled access and human review. They should balance the organisational context needed for actionable findings against unnecessary data exposure. Architectures allowing firms to switch between models and tools may preserve flexibility and reduce supplier dependence.
Firms should test their organisational readiness for a greater volume and complexity of vulnerability findings. This includes identifying bottlenecks across validation, engineering, patch testing and change implementation; distinguishing plausible outputs from genuinely exploitable vulnerabilities; and prioritising potential attack paths by their impact on important business services. Firms should consider whether urgent remediation can be undertaken while managing operational and change risks.
Finally, senior leaders should have clear accountability and visibility of cyber risk, remediation capacity and operational resilience. Firms should understand the assets, dependencies and suppliers supporting important business services, assess providers’ preparedness for AI-enabled vulnerability discovery, and test severe but plausible scenarios involving simultaneous disruption across firms or shared technology infrastructure.
“Frontier AI presents a significant opportunity for firms to strengthen their cyber defences and resilience. Realising that potential safely and efficiently will require firms to adapt their controls and build the capacity to detect, prioritise and respond to vulnerabilities at greater speed and scale without overwhelming existing processes.”
Alex Petsopoulos
Partner, PwC Cyber, Data and Tech Risk Lead
Next steps
By the end of 2026, the FCA plans to publish examples of good and poor AI practice, while the BoE intends to consult on the management of information and communication technology and cyber risks. UK authorities will continue engaging with industry through initiatives such as the FAISF.
Contacts
Hugo Rousseau
Follow us
