This publication is issued jointly by the Prudential Regulation Authority (PRA), the Financial Conduct Authority (FCA), and the Bank of England (the Bank) (collectively, the regulators). Where it refers to ‘we’, ‘us’, or ‘our’, this means the regulators collectively. In this publication, ‘firm’ is used in the same way it is used in our supervisory statement (SS) and Final Guidance (FG) (throughout ‘SS/FG’) on cyber risk and resilience management and includes financial market infrastructures (FMIs).
1: Introduction
1.1 This publication highlights effective practices that firms can adopt to manage cyber risk and resilience in a world where artificial intelligence (AI) is rapidly creating new cyber risks and amplifying existing ones, while simultaneously enabling firms to strengthen their cyber defences.
1.2 The effective practices are complementary to, but separate from, our SS/FG on cyber risk and resilience management.footnote [1] The regulators’ expectations in the cyber risk and resilience management SS/FG apply regardless of the technology firms use. The effective practices provide more detail on areas relevant to AI. They are, however, not part of the regulators’ formal expectations.
1.3 The effective practices build on relevant publications by the regulators, HM Treasury, the National Cyber Security Centre (NCSC), non-UK authorities, international standard-setting bodies, and industry. They have also been informed by our engagement with firms.
1.4 AI has both cyber defensive and offensive capabilities. The effective practices highlight areas where firms can use AI to strengthen their cyber defences. However, it is important not to over-rely on AI. Appropriate human oversight remains essential, especially for key decisions.
1.5 The effective practices use the Organisation for Economic Co-operation and Development (OECD) definition of an AI system.footnote [2] They avoid terms such as ‘frontier AI’, which can change meaning or become ambiguous as technology evolves. Nonetheless, the effective practices are particularly relevant for AI systems whose autonomy and capabilities create high cybersecurity risks and, in many cases, opportunities to strengthen firms’ cyber defences.
1.6 The effective practices are intended to be used proportionately. Firms can adopt the practices in a manner appropriate to their size, complexity and risk profile. In particular, they may consider:
- their reliance on AI for important business services and critical operations; and
- the risks of AI systems they rely on, including their autonomy, their level of access to the firms’ assets, and their ability to interact with the external environment.
Summary
1.7 Recent developments in AI are having a material impact on firms’ cyber risk and resilience by significantly reducing the cost, skills and time required for threat actors to discover and exploit vulnerabilities. Advanced AI systems can analyse source code, identify vulnerabilities and discover attack paths at great speed – allowing threat actors to exploit vulnerabilities within hours or days rather than weeks or months. While AI is increasing the scale and speed at which vulnerabilities can be identified and exploited and reducing technical barriers to entry, humans remain central to advanced cyber threat activity.footnote [3]
1.8 Firms’ traditional cyber defences may struggle to cope with this new landscape. As noted by the UK AI Security Institute (AISI):
- AI capabilities are advancing rapidly across all tested domains. Performance in some areas is doubling every eight months, and expert baselines are being surpassed rapidly;
- safeguards are improving but vulnerabilities remain;
- some capabilities required for AI models to evade human control are emerging; and
- the performance gap between open and closed source models has narrowed over the past two years.
1.9 Due to the increase in the number of vulnerabilities being discovered and the speed at which they are being identified, firms need to evolve from traditional vulnerability management, where discovery has historically been the main bottleneck, towards continuous exposure management. The key challenges that firms face in this evolution are to prioritise, validate, remediate and recover from threats at machine speed.
1.10 This evolution towards exposure management means that cyber risk is no longer a purely technical issue. It is a core business risk and leadership responsibility. Effective practice for boards and senior management is not only to ensure that the firm has appropriate cyber controls (design) but also to obtain assurance that those controls will perform appropriately during incidents (effectiveness). This may involve using AI to strengthen cyber defences and reassessing trade-offs; for example, remediating a critical vulnerability quickly even if doing so could disrupt a service.
1.11 Effective governance, including board oversight and senior management accountability, remains critical. However, to be effective, the nature and speed of reporting to and decision-making by boards and senior management need to evolve. Strategic plans can help fix a growing number of vulnerabilities quickly, supported by appropriate governance and resources. In addition to faster decision-making, effective practice may require that firms also increase the speed at which they identify affected assets, test remediation, secure a change window, coordinate with third parties, deploy remediation and verify its success (known as execution speed).
1.12 AI-related cyber risks could threaten financial stability if they affect shared resources used by many firms. This includes critical third parties, and common data, infrastructure and software (including open source). Cooperation across borders and sectors, with collective defences operating at machine speed, is essential to manage these risks.
2: From vulnerability to exposure management
2.1 Traditionally, firms dealt with cyber risks through vulnerability management – periodically identifying and addressing vulnerabilities. To cope with the new speed and volume of potential attacks, effective practice means firms evolving towards exposure management. This involves identifying exposures (any misconfigurations, security gaps or weaknesses that a threat actor could exploit) and managing them systematically. Firms can use AI to enhance their exposure management capabilities but should bear in mind that advanced AI systems are non-deterministic: they may produce different results across repeated assessments, making validation of their outputs important. Moreover, the respective capabilities of firms to use AI for cyber defence and of threat actors to use AI for cyber-attacks are not always symmetrical. For instance, effectively managing cyber risks requires firms to secure financial and non-financial resources and adopt appropriate governance. Conversely, threat actors typically just need to focus on identifying and exploiting vulnerabilities while avoiding detection.
Identifying exposures
2.2 The first step in managing exposures effectively is to identify them. This involves determining, mapping and understanding which assets a threat actor could exploit, and the connections and dependencies between those assets.
2.3 Exposures can arise in:
- AI systems, including those available with open weights;
- application programming interfaces (APIs) and other model endpoints;
- cloud services;
- external connectors (components that link an AI system with external data sources, tools or APIs);
- hardware and software;
- identities (human and non-human, such as AI agents);
- material third-party dependencies, including via:
- AI embedded in third-party services;
- key Nth party and supply dependencies;
- third-party data;
- infrastructure dependencies such as hardware, data centres, telecommunications, power, water etc; and
- open-source and open-weight dependencies, some of which may be widely embedded in other software and not immediately visible.
2.4 Firms can use existing sources to help identify and document exposures, including:
- AI and model inventories;
- operational resilience maps and self-assessments;
- outsourcing and third-party registers; and
- software bills of materials.
2.5 Firms may need to carefully manage gaps, inconsistencies and silos in the data obtained from these sources. Firms can use dynamic asset identification tools to automate parts of the identification process and identify potential blind spots.
2.6 Firms may prioritise the identification of exposures linked to their important business services or critical operations.
2.7 Identifying and documenting exposures can also help firms address the unauthorised use of applications (‘shadow IT’), including those developed using AI, which are a significant source of riskfootnote [4].
Reducing the attack surface, including managing legacy assets
2.8 Effective practice involves firms controlling and reducing their cyber-attack surface in all circumstances, but particularly in relation to AI-powered attacks. These attacks disproportionately threaten exposed, legacy, unmanaged and poorly managed assets. These assets are a time-sensitive business liability for firms. Once a firm has identified its AI exposures, the next step is to develop a comprehensive strategy to progressively modernise its IT.
2.9 Firms can control and reduce their attack surface by:
- decommissioning, disabling, isolating, or replacing end-of-life or end-of-support assets;
- eliminating unnecessary internet exposures;
- closing or disabling inactive or unnecessary access points;
- deleting orphaned or unused accounts and stale identities;
- eliminating redundant privileged accounts;
- limiting or revoking excessive or unnecessary permissions;
- fixing weak settings (for example, changing factory default administrator usernames and passwords in assets provided by third parties);
- segmenting or micro-segmenting their most valuable assets; and
- using web application firewalls and edge security tools to strengthen their perimeter.footnote [5]
Continuous exposure monitoring
2.10 Effective exposure management involves shifting from periodic, routine, vulnerability identification to continuous exposure monitoring. Continuous exposure monitoring involves comprehensively covering externally exposed assets, configurations, identities and systems, including connections to third parties. Firms may begin by prioritising external-facing attack surfaces, technologies on the perimeter and then work inwards to cover on-premises and cloud environments. It is important to include AI-specific attack surfaces such as embeddings (numerical representations of data that help AI systems capture relationships between concepts), memory modules and tool chains.
2.11 Firms can use attack surface management tools to automate their monitoring and help them discover vulnerabilities, although it is important to retain appropriate human oversight, scrutiny and validation of any findings made by these tools.
3: Risk-based prioritisation
3.1 To manage exposures effectively, firms need to prioritise assets and vulnerabilities based on the risks they pose. For example:
- vulnerabilities that threat actors are already actively exploiting (as reported in public databases and other sources of threat intelligence);footnote [6]
- assets supporting important business services or critical operations;
- material third-party services;
- externally exposed assets (those connected to the internet); and
- assets at risk of becoming obsolete (ie at their end-of-life or end-of-support stages).
3.2 These categories are not mutually exclusive. Firms may need to consider them in combination. This can be facilitated by grouping and addressing vulnerabilities by type rather than individually.
3.3 To be effective, firms’ prioritisation criteria for vulnerabilities may need to evolve beyond individual metrics and consider:
- how easy it would be for threat actors to exploit a vulnerability;
- potential attack paths (including across connected or dependent assets);
- the potential impact of a vulnerability being exploited; and
- the potential aggregate impact of multiple individual low-level vulnerabilities being exploited – for instance, if AI systems chain these vulnerabilities.
3.4 Firms can set risk-based remediation targets for priority vulnerabilities, including emergency remediation for those that are most critical or time-sensitive. For example, they might seek to recover important business services in a set number of hours or days. At the same time, firms need to consider how to retain the capacity to manage spikes in general remediation activity.
3.5 To be effective, firms need to treat remediation speed and vulnerability backlogs as operational risks.
3.6 Automating the identification, triage, prioritisation and remediation of vulnerabilities (or aspects thereof) can be an important enabler of exposure management. Where automatic patching is feasible, its use is strongly encouraged, particularly ‘hot patching’ (automatic patching without service disruption).footnote [7] Automation is most effective when firms properly assess the risks and overlay human oversight and safeguards. This allows firms to augment, but not replace, human judgment for key decisions.
4: Harness engineering
4.1 A harness is the collection of tools, workflows, controls, data and operating environments that sit around advanced AI systems and shape how they are used. In cyber defence, a harness defines what information the AI system produces so that people can validate and act on it. The harness can determine:
- what information is provided to the system;
- what tasks it performs;
- how its outputs are checked; and
- how its findings are given to cybersecurity and engineering teams.
Effective harness engineering is key to using AI for cyber defence.footnote [8] The number of vulnerabilities a firm discovers and remediates may depend as much on the harness as on the underlying AI.
4.2 It is important for firms to recognise that access to advanced AI systems does not automatically create or enhance effective cyber defence. Instead, effective practice will likely involve designing a well-structured harness that can turn AI capabilities into outputs that are usable, sufficiently explainable and actionable. A well-designed harness includes:
- effective workflow design;
- controlled access to tools;
- contextual information provided to the AI system;
- validation steps;
- integration with existing engineering and security platforms; and
- meaningful human review (including targeted retesting of the AI system’s findings).
While full explainability of AI outputs might not be technically possible, the harness should ensure the firm can produce evidence of how it has validated the outputs and who is accountable for them.
4.3 Harness approaches could include internally developed tools, vendor-provided platforms, managed third-party offerings and open-source components. There may not be a single approach that provides a complete solution, so firms may combine them to balance flexibility, transparency, supportability and integration with their existing processes.
4.4 Firms may need to consider ways to:
- coordinate multiple harnesses;
- route activity to specialist tools or models;
- compare capabilities;
- improve signal-to-noise identification (the ability to distinguish genuine threats from false positives and normal behaviour); and
- reduce dependency on a single model or vendor.
4.5 A modular architecture can help firms adapt as technology evolves. However, this may have limitations: different harness components may not be easily interchangeable, harness behaviour can be highly system-specific, and changes may require re-validation.
4.6 To get usable cyber defence outputs from AI systems, firms will likely need to provide them with sufficient context about their organisational model and structure, while avoiding unnecessarily exposing sensitive information. This context may include firms’ approaches to classification, supplier assurance, data access and exclusions. Where appropriate, firms may use isolated, sandboxed environments.
4.7 Firms may design harnesses that allow them to increase their capacity to validate, prioritise and remediate issues at scale. Scaling the use of advanced AI systems depends as much on people, processes and remediation capacity as on the AI capability itself.
4.8 When measuring the effectiveness of their harnesses, it is important for firms to consider not only the volume of vulnerabilities identified but also the number of vulnerabilities that were disclosed, mitigated or remediated in ways that meaningfully reduced risks.
5: Mitigating and compensating controls
5.1 Effective practice in managing exposures involves firms balancing the need to address vulnerabilities quickly with the need to maintain service continuity. In practice, there may be trade-offs between the speed of remediation and an increased risk of operational disruption. Managing these trade-offs needs effective controls, which may include:
- involving key stakeholders in vulnerability remediation – potentially including relevant business areas, communications and control functions, technology specialists and material third parties;
- applying clear criteria for patching and other vulnerability management measures;
- testing remediation measures (and automating where possible);
- having fallback plans and procedures; and
- linking remedies to response and recovery.
5.2 Likewise, where patches or other remediation are not available or cannot be applied immediately, firms can implement compensating controls (which may be temporary). These might include:
- restricting access to assets where vulnerabilities have been identified;
- isolating assets and disabling services;
- filtering traffic; and
- virtual patching.footnote [9]
5.3 In addition to compensating controls, it is important for firms to document instances where vulnerabilities cannot be remediated immediately, assign an accountable owner for each instance, and agree a dated remediation plan.
5.4 Reports to the firm’s board and senior management may highlight areas where the firm remains exposed, residual risks (including those that were accepted), trade-offs and compensating controls.
6: Third-party, open-source and supply chain risk
6.1 To effectively manage exposures, firms can document their third-party and supply chain dependencies. This includes dependencies on third parties they have a direct contractual relationship with, as well as on:
- open-source and open-weight tools;
- material third parties’ key Nth party and supply chain dependencies; and
- as far as possible, AI embedded in non-AI material third-party services and products.
6.2 Firms may need to invest in capabilities to get this granular visibility. For instance, software composition analysis tools can help identify open-source dependencies that may not be immediately visible to firms.
6.3 Focusing on material third-party dependencies, single points of failure and connections between third-party services and third parties can facilitate risk-based prioritisation.
6.4 Firms may depend on third parties to remediate vulnerabilities in the services they supply to them. Third parties with access to restricted AI systems can also play a key role in spreading the benefits of these systems for cyber defence, and for remediating vulnerabilities in widely used software. It is important for firms to understand and assess the efficiency and speed of their third parties’ processes for identifying and fixing vulnerabilities. One aspect firms may need to consider is the capacity of smaller but material third parties to cope with a high volume of discovered vulnerabilities. Firms, industry groups and third parties may consider working collaboratively to address this.
6.5 Firms may need to review and update contracts with material third parties to address AI risks. Revisions might include:
- disclosure of embedded AI in material third-party services;
- service levels for vulnerability remediation – for example, frequency and volume of patching; and
- prompt notification of changes to AI systems that may impact the third party’s cyber risk profile (such as increased or reduced safeguards).
6.6 Open-source and open-weight dependencies can have benefits such as a decentralised community of experts that can collaborate to quickly remediate vulnerabilities. However, they also have corresponding risks. In particular, firms do not have direct recourse to a specific third-party and may need to implement compensating measures. For instance:
- updating their assessment and approval processes and standards for open-source and open-weight tools; and
- as far as possible, reducing their dependency on unsupported open-source or open-weight tools to reduce their attack surface.
7: Authentication and identity and access management
7.1 Robust authentication and identity and access management, including access controls, multi-factor authentication and ongoing monitoring are key to effective exposure management.
7.2 To protect against AI-enabled social engineering attacks, such as sophisticated deepfakes and impersonation, firms may use tools such as AI behavioural analysis and phishing-resistant multi-factor authentication. Placing an independent authentication layer in front of sensitive internet-facing systems can also be useful.
7.3 Where firms are granted access to third-party AI systems with advanced capabilities on a pre-release or restricted basis for the purposes of cyber defence, they may review and, if appropriate, strengthen their identity and access management to mitigate the risk of unauthorised users accessing those systems, or those systems gaining unauthorised access to the firm’s assets. This includes managing the risk of misuse by malicious insiders.
Secure software development
7.4 In line with our operational resilience framework, firms need to assume that they will suffer disruption at some point (including from AI), and design their systems to contain, withstand, respond to and recover from sustained, machine-speed attacks. They may consider:
- adopting secure-by-design practices;
- integrating security into every phase of software development;
- moving security testing to the earliest phase of software development (‘shift left’);
- performing threat modelling (incorporating AI-specific threats, such as adversarial prompts and jailbreaking);
- embedding automated security testing into software development;
- scrutinising AI-generated as well as human-generated code;
- using AI code review tools;
- continuously scanning source code, infrastructure-as-code and deployment pipelines;
- protecting the confidentiality of their source code;
- increasing the use of memory-safe technologies and secure coding;
- using AI secure-development tools to stop vulnerabilities entering production;
- integrating remediation into software deployment pipelines; and
- designing assets and systems to degrade gracefully over time.
Design for containment and resilience
7.5 Firms might take steps to ensure that they can contain attacks by:
- applying ‘zero trust’ principles;
- implementing strong segmentation and (if applicable) micro-segmentation;
- limiting attackers’ lateral movement (for instance, through deception techniques);
- enforcing ‘least privilege’;
- adopting just-in-time privileged access and monitoring privileged actions continuously;
- using exploit-prevention technology;
- using real-time monitoring and centralised logging;
- building predefined emergency containment pathways; and
- designing systems to rapidly isolate compromised components.
8: Response and recovery from AI-enabled cyber-attacks
8.1 Firms may need to update their frameworks to ensure they can detect, respond and recover effectively from attacks involving rapid exploitation of newly discovered vulnerabilities and AI-related cyber-attacks. The latter may include:
- cyber-attacks aimed at AI systems, such as:
- data or model poisoning attacks (which can take a long time to cause detectable harm);
- malicious input attacks (extraction, evasion, jailbreaking, prompt injections, misuse of retrieval-augmented generation pipelines); and
- cyber-attacks by highly autonomous AI agents.
8.2 Continuous exposure monitoring, as described above, can help detect incidents and anomalous performance by AI systems. To detect AI-specific incidents, monitoring may need to cover prompt patterns, output anomalies and system behaviour. Comparing logs from LLM APIs, tool executions and chat sessions can be an effective detection mechanism. Firms may also use LLM-as-a-judge tools (which use large language models to evaluate text, including other AI-generated text) to detect anomalous behaviour and potential incidents.
8.3 Where appropriate, firms may develop AI-specific incident response playbooks or enhance their existing playbooks to ensure that issues such as escalation thresholds, emergency decision-making authorities and processes, evidence requirements, and authority to isolate, reduce or take services offline, are clear and documented before an incident arises.
8.4 AI can help firms automate and optimise their cyber response and recovery by:
- providing data to help them triage and prioritise issues (such as data on affected assets, compensating controls, dependencies, exposures, exploitability and potential impact);
- automating routine tasks to free up staff and management time for making decisions requiring human judgment;
- making forensic analysis and evidence-gathering faster and more consistent; and
- helping teams coordinate (such as engineering, operations and security).
8.5 In an AI-specific context, recovery activities may need to include restoring compromised AI services, rolling back fine-tuned models or agent states, or redacting unsafe memory or embeddings. In incidents involving compromised data, recovery may also involve updating retrieval-augmented generation sources and retraining the AI system with uncompromised data.
8.6 Automation can be particularly useful in the response and containment stage, where firms can use predefined, tested playbooks to:
- carry out forensic examinations;
- hunt for threats;
- isolate compromised assets and revoke access;
- execute early defensive actions at speed; and
- help staff in security operations centres.
9: Recovery
9.1 Firms may consider updating their recovery plans to ensure they can recover from AI-related cyber-attacks. Recovery plans may need to consider:
- a greater likelihood of zero-day attacks;
- faster impact and propagation;
- simultaneous impact on multiple assets (for example, through chained vulnerabilities); and
- AI agents gaining unauthorised access to recovery environments.
9.2 Recovery options may include (but are not limited to) manual fallbacks and immutable backups for important business services and critical operations.
9.3 Where possible, validating the integrity of services suspended following an incident before restoring them enables effective recovery and can prevent second-order incidents. There may, however, be situations where restoring a service quickly is less of a risk than waiting for a full validation.
10: Human oversight
10.1 Effective human oversight is a key cross-cutting element when firms either: (i) use AI to automate aspects of their cyber defence; or (ii) use highly autonomous, capable AI systems in their business or operations. Effective human oversight is meaningful, risk-based and capable of changing outcomes. It improves decisions, prevents harm and enables errors and their impacts to be addressed, rectified or reversed. People overseeing AI need sufficient understanding of its capabilities and limitations, together with the authority, time, information and incentives to challenge, override or stop it if warranted.
10.2 The form and intensity of oversight needs to reflect the AI use case’s materiality, risk, autonomy, complexity and explainability. Human-in-the-loop approval may be appropriate where errors could have serious or irreversible consequences; human-on-the-loop monitoring may suit lower-risk activity where intervention is needed only when defined triggers are met. Human-in-command arrangements can set autonomy, guardrails and escalation routes, particularly for highly autonomous AI. Contestability and kill switches (which already feature in algorithmic trading)footnote [10] can enable decisions to be appealed and systems paused, constrained or shut down safely.
10.3 Clear, explicit accountability across the lifecycle and appropriate training are key to effective human oversight. Oversight quality may itself need to be assessed, including whether human reviewers are critically assessing AI outputs rather than accepting them without sufficient scrutiny (known as ‘automation bias’) and exercising informed judgement.
10.4 Effective oversight also depends on observability and feedback. Organisations may need to continuously monitor AI behaviour, retain logs of actions and intermediate decisions, investigate anomalies, track human overrides and use lessons from interventions to improve both the system and the oversight process. Human approval for high-value, sensitive or potentially harmful actions based on clear triggers is another key control.
10.5 As firms scale their use of highly autonomous AI systems, human oversight reinforced by deterministic technical controls may be appropriate. These include least-privilege access, secure operating environments, limits on tools and external connections, staged increases in autonomy, prohibited actions and controls at each step of a task. Regular spot checks, testing, red-teaming and resilience exercises should confirm that systems escalate, stop, hand over or return to a safe state when necessary. Higher-risk systems warrant layered controls, more intensive review and clear accountability for outcomes.footnote [11]
11: Assurance
11.1 Firms’ assurance activities may need to evolve to address AI-related cyber risks by:
- testing severe but plausible cyber disruption scenarios involving AI – for example, hallucinations or prompt injections (these scenarios can be informed by threat intelligence and industry guidance); and
- adjusting the severity of scenarios to consider risks from AI systems with high autonomy and capability – for instance, by increasing the speed at which disruption crystallises.
11.2 Assurance activities are most valuable when they assess not just a firm’s technical capabilities, such as the speed at which it can patch a vulnerability, but also the effectiveness and speed of leadership decisions (including reprioritisation and resource allocation) and the firm’s continued ability to meet its contractual, customer and regulatory obligations under stress.
11.3 Future industry and sector-wide exercises may need to consider scenarios involving AI, such as high-speed high-volume attacks, broad compromise through zero-day or chained vulnerabilities, autonomous attacks by AI agents, and disruption to material third parties and key infrastructure.
12: Sector collaboration
12.1 Collaboration and engagement is increasingly important, including through collective action and incident response frameworks such as the Cross-Market Operational Resilience Group (CMORG)footnote [12] and the Frontier AI Information Sharing Forum (FAISF). These fora can help firms respond to machine-speed attacks in a coordinated way and mitigate potential systemic risks.footnote [13] Among other activities, collective action enables firms to:
- develop mechanisms for collaboration before vulnerabilities are disclosed;
- participate in sector-wide initiatives, such as vulnerability clearinghouses (public and private sector led);
- share the costs of vulnerability remediation activities like validation or patching;
- identify and disseminate best practices; and
- strengthen the sector’s collective ability to respond to incidents.
13: Conclusion
13.1 Developments in AI are having a transformative effect on cyber risk. Cyber hygiene, effective governance and traditional technical controls will continue to remain essential. The main change is likely to be the unprecedented speed and volume in which cyber risks emerge and crystalise. This will require an adaptive mindset by firms, and an evolution in traditional cyber risk and resilience management across the sector.
References
Financial authority publications
Bank of England, Financial Conduct Authority, HMT:
- The Bank, FCA and HM Treasury joint statement on Frontier AI models and cyber resilience | Bank of England – the UK's central bank (May 2026)
- Frontier AI Information Sharing Forum (FAISF) | Bank of England – the UK's central bank
- Frontier AI: Harness engineering | Bank of England – the UK's central bank
- Frontier AI and cyber resilience | FCA
- AI in financial services: good practice and areas for improvement | FCA
Financial Stability Board (FSB):
- The Financial Stability Implications of Artificial Intelligence - Financial Stability Board (2024)
- Monitoring Adoption of Artificial Intelligence and Related Vulnerabilities in the Financial Sector - Financial Stability Board (2025)
- Sound Practices for Responsible Adoption of Artificial Intelligence (AI): Consultation report - Financial Stability Board (October 2026)
Bank for International Settlements (BIS):
- A Mythos moment? Frontier AI and cyber risk (July 2026)
- When machines attack: frontier AI cyber threats and policy responses in the financial sector (September 2026)
European Central Bank (ECB):
- Addressing AI-enabled cybersecurity threats (July 2026)
European Supervisory Authorities (ESAs):
- ESA Statement on frontier AI models - European Insurance and Occupational Pensions Authority (July 2026)
European Systemic Risk Board (ESRB):
G7 Cyber Expert Group:
Australian Prudential Regulation Authority:
- APRA Letter to Industry on Artificial Intelligence (AI) | APRA (Apr 2026)
- Insights from the APRA-ASIC Industry Roundtables | APRA (Aug 2026)
Canadian Office of the Superintendent for Financial Institutions (OSFI):
- Frontier Artificial Intelligence: Implications for Technology, Cyber Security, and Operational Resilience - Office of the Superintendent of Financial Institutions (Apr 2026)
- Generative and Agentic Artificial Intelligence: Implications for Technology, Cyber Security, and Operational Resilience - Office of the Superintendent of Financial Institutions (July 2026)
Hong Kong Monetary Authority:
Japan FSA:
US NY DFS:
- Measures Regulated Entities Should Consider in a Heightened Cybersecurity Threat Environment (May 2026)
- Heightened Cybersecurity Risks Associated with Frontier AI Models (May 2026)
UK Government and NCSC sources
National Cyber Security Centre (NCSC):
- Frontier AI: What You Need to Know
- Retaining Defensive Advantage in the Age of Frontier AI Cyber Capabilities (Apr 2026)
- Preparing for a 'vulnerability patch wave' (May 2026)
- Managing the cyber risk of agentic AI | National Cyber Security Centre (Aug 2026)
- The hidden risks of shadow AI | National Cyber Security Centre (Sept 2026)
- Impact of AI on cyber threat from now to 2027 | National Cyber Security Centre
Five Eyes:
UK Government AI Safety Institute (formerly AI Safety Institute):
- Various publications on frontier AI safety, model evaluations and cyber capabilities.
Industry and technical guidance
Cross-Market Operational Resilience Group:
The SS/FG is now subject to consultation.
An AI system is a machine-based system that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. Different AI systems vary in their levels of autonomy and adaptiveness after deployment. See What is AI? Can you make a clear distinction between AI and non-AI systems?
Also see The Bank, FCA and HM Treasury joint statement on Frontier AI models and cyber resilience.
See The 'vibe coding spectrum' approach to AI-assisted software development and The hidden risks of shadow AI.
Edge tools are hardware, software, and cloud-based solutions used to protect data, devices, and workloads at the outer boundaries of a network where users and endpoints connect, rather than at a centralised data centre.
For example CVE: Common Vulnerabilities and Exposures, and National Vulnerability Database.
Frontier AI: Harness engineering and Frontier AI and cyber resilience.
Virtual patching is a security policy enforcement layer which does not fix a known vulnerability, but mitigates its exploitation, for example through monitoring of exploitation attempts and intercepts such attempts.
UK MiFID RTS 6 Article 12 - Kill Functionality, FCA Handbook – MAR 7A.3 Requirements for algorithmic trading, and PRA SS5/18 – Algorithmic trading.
Managing the cyber risk of agentic AI | National Cyber Security Centre, and Careful adoption of agentic AI services | CGA.
Relevant CMORG publications include: AI Risk Mitigation Plan | Cross Market Operational Resilience Group; Firm Guidance for Frontier AI | Cross Market Operational Resilience Group; and Frontier AI: Industry Preparedness Summary | Cross Market Operational Resilience Group.