Effective practices: Cyber risk and resilience management relating to artificial intelligence (AI)
AI Analysis
On 19 November 2026, the PRA, FCA and Bank of England published non-binding effective practices for managing cyber risk and resilience associated with AI. The central market message is that firms should move from periodic vulnerability management to continuous exposure management, with faster remediation, stronger mapping of AI and third-party dependencies, meaningful human oversight, and board assurance that controls work during incidents.
Key dates
- 2026-11-19
- The PRA, FCA and Bank of England published the AI cyber risk and resilience effective practices. The practices are not subject to formal consultation and have no stated implementation deadline.
- 2027-02-26 Deadline
- The separate joint consultation on cyber risk and resilience management, CP17/26 and GC26/9, closes. This deadline does not apply to the AI effective practices themselves.
Suggested considerations
- Firms should consider treating the publication as a supervisory signal and benchmarking current cyber, AI governance and operational-resilience arrangements against its practices, while recognising that the practices are not formal requirements.
- Compliance and technology teams may wish to reconcile AI inventories, model inventories, operational-resilience maps, outsourcing and third-party registers, software bills of materials and identity inventories, including open-source, open-weight, cloud, API, external-connector and nth-party dependencies.
- Firms should consider prioritising exposure discovery and continuous monitoring for assets supporting important business services or critical operations, externally exposed systems, actively exploited vulnerabilities, material suppliers and end-of-life or end-of-support assets.
- Firms may wish to establish documented, risk-based remediation targets, emergency remediation procedures, compensating controls, accountable owners and dated remediation plans where immediate remediation is not feasible; remediation speed and vulnerability backlogs should be considered within operational-risk reporting.
- Technology and security teams should consider reducing attack surfaces by decommissioning unsupported assets, eliminating unnecessary internet exposure, removing stale identities and redundant privileged accounts, restricting permissions, segmenting valuable assets and correcting weak default settings.
- Where AI is used for cyber defence, firms should consider documenting the harness architecture, permitted tools and data, validation and retesting controls, human approval points, audit evidence, model or vendor dependencies, false-positive handling and change-revalidation procedures.
- Boards and senior management may wish to revise cyber reporting so that it supports faster decisions and provides assurance about control effectiveness during incidents, including remediation execution speed, third-party coordination and recovery of important business services.
- Firms should consider exercising AI-enabled attack scenarios and testing whether governance, escalation, decision rights, third-party coordination and technical recovery arrangements operate at the speed required by machine-assisted attacks.
What changed
The publication does not create new formal rules or supervisory expectations and is separate from the regulators' cyber risk and resilience supervisory statement and FCA final guidance. It provides proportionate, technology-agnostic practices for firms and FMIs, including mapping AI systems, model endpoints, identities, cloud services, external connectors, open-source and open-weight dependencies, and material third-party and nth-party dependencies. It encourages continuous exposure monitoring, attack-surface reduction, risk-based prioritisation using active exploitation, attack paths and business-service impact, emergency remediation, automatic patching and hot patching where feasible, and treatment of remediation speed and vulnerability backlogs as operational risks. Firms using AI for c
Compliance impact
The immediate legal impact is limited because the publication expressly states that the practices are not part of the regulators' formal expectations. Nevertheless, the regulators' emphasis on continuous exposure management, rapid remediation, board assurance and AI-related third-party dependencies is a significant supervisory signal, and weaknesses may become relevant when firms demonstrate compl
Who is affected
Related regulations
References
AI-generated analysis. May contain errors or omissions — verify with the original PRA source before acting. Full disclaimer.
What the PRA said
This joint PRA, FCA and Bank publication highlights effective practices to manage cyber resilience and risks relating to artificial intelligence (AI) for firms and financial market infrastructures.
Published by PRA . Read the full notice at the source for the authoritative text.