Summary
- The Bank of England, PRA and FCA have published effective practices addressing both AI-assisted attacks and security weaknesses in enterprise AI systems.
- Recommendations cover continuous exposure monitoring, faster remediation, secure development, agent permissions and recovery from AI-specific incidents.
- Human oversight must be capable of changing outcomes, supported by technical restrictions and clear accountability for automated actions.
UK financial regulators are encouraging institutions to move beyond periodic vulnerability assessments and strengthen their ability to identify, contain and recover from attacks accelerated by artificial intelligence.
The Bank of England, Prudential Regulation Authority and Financial Conduct Authority published effective practices for managing AI-related cyber risk on 19 November 2026, addressing both the use of AI by attackers and the security of increasingly autonomous systems operated by financial firms.
The publication accompanies a wider consultation on cyber resilience management but is not part of the regulators' formal supervisory expectations. Institutions can adopt the suggested practices proportionately, taking account of their size, complexity, operations and reliance on AI.
The regulators identify faster vulnerability discovery and exploitation as major changes in the threat environment. Advanced AI systems can analyse code, identify weaknesses and investigate potential attack paths more quickly, placing additional pressure on organisations whose remediation processes depend on fixed maintenance cycles.
Exposure management extends beyond known vulnerabilities
Traditional vulnerability management concentrates on identifying software weaknesses and arranging corrective action. The regulators are encouraging a broader approach that also considers misconfigurations, unnecessary internet exposure, excessive permissions, unmanaged assets and relationships with external providers.
Continuous exposure monitoring can establish where those conditions exist and how they affect important business services. Rather than treating every finding independently, organisations can examine whether several weaknesses create a practical route towards sensitive infrastructure.
Existing information can support the work, including asset inventories, operational resilience maps, third-party registers and software bills of materials.
AI introduces further dependencies. Models may interact with retrieval systems, memory stores, application programming interfaces and external tools, each creating possible paths through which malicious information or unauthorised instructions could affect the system.
The guidance therefore recommends including embeddings, memory modules and agent tool chains within exposure assessments.
AI tools can assist with monitoring and prioritisation, but their output requires validation. Because model behaviour can vary across repeated assessments, an automatically generated finding should not be treated as conclusive evidence that a particular system is exploitable.
Organisations also need the operational capacity to act on findings. Detecting a critical weakness quickly provides limited protection if identifying the system owner, testing a correction and obtaining permission to deploy it takes considerably longer.
Autonomous agents require enforceable boundaries
Financial firms are increasingly considering AI agents that can retrieve information, interact with business applications and carry out tasks with limited direct supervision.
Such systems create risks where their permissions extend beyond the activity they are intended to perform. An agent connected to administrative tools may be able to make changes that have consequences for several applications, even when it receives instructions through an apparently routine user request.
The regulators recommend least-privilege access, controlled operating environments, restrictions on permitted tools and external connections, and limits on actions with substantial consequences.
They also identify the value of technical controls that prevent an agent from bypassing restrictions, alongside arrangements for escalating decisions that require human approval.
Meaningful oversight should reflect the risks involved. A high-impact operation may require approval before execution, while lower-risk activity may be monitored with intervention when predefined conditions arise.
In either case, the person supervising the system needs sufficient information, authority and time to challenge its behaviour.
Repeated approval requests can create their own weakness if employees accept decisions without examining them. The regulators specifically identify automation bias as a concern, recommending checks that establish whether human reviewers are exercising informed judgement.
Incident response must account for AI-specific failures
The guidance addresses attacks that manipulate AI systems directly, including prompt injection, model or data poisoning, and malicious instructions introduced through retrieved information.
These incidents may differ from conventional malware attacks. A model could remain accessible while producing unsafe results, using inappropriate tools or relying on information that has been deliberately manipulated.
Investigators may therefore need records of prompts, model outputs, tool execution and intermediate actions, alongside conventional application and infrastructure logs.
Where an AI system has been compromised, recovery might involve restoring a model, rolling back an agent's operating state or removing unsafe information from memory and retrieval systems.
Those activities require coordination between security, engineering and operational teams, particularly where an automated system supports an important business service.
AI can also assist incident response by collecting evidence, correlating information and performing routine investigative tasks. The regulators nevertheless emphasise retaining human responsibility for decisions involving material operational consequences.
Shared technology creates concentrated exposure
The publication identifies common software, cloud infrastructure, open-source components and critical third-party providers as possible sources of simultaneous exposure across the financial sector.
Where several institutions rely on the same technology, a weakness or compromise may affect multiple organisations before each has completed its own investigation.
That possibility increases the importance of reliable supplier information and coordinated response arrangements, particularly when attacks develop faster than established communication or change approval procedures.
The regulators also encourage resilience exercises involving severe but plausible AI-related disruptions, including situations where malicious activity develops rapidly or an automated system produces unreliable results.
Those exercises can test whether an institution can restrict a compromised agent, recover affected information and maintain essential operations without depending on the continued correctness of the AI system itself.
The new effective practices do not mandate particular security products or technical architectures. They establish a framework for assessing how existing cybersecurity responsibilities apply when AI increases both the pace of attacks and the autonomy of systems operating within financial institutions.





