Britain’s financial regulators have opened a joint consultation that does something more useful than add another cyber checklist: it tries to define what good cyber resilience looks like when AI can speed up reconnaissance and exploitation, and when the most important weaknesses often sit outside the firm’s own data centre. In GC26/9 and CP17/26, the FCA, PRA and Bank of England say the draft guidance clarifies existing expectations rather than creating new rules. But for banks, insurers, exchanges, payment firms and key suppliers, the operating consequences are immediate: boards will be expected to show not just that vulnerabilities were logged, but that exposures were found, prioritised, contained, fixed, tested and recoverable.
The practical question for firms is not whether they can buy another detection tool. It is whether they can prove, under pressure, that they can identify, contain, recover from and learn from an AI-accelerated attack when the weak point may be a cloud service, software dependency, model endpoint, external connector or shared provider they do not fully control.
From patch cycles to exposure management
That is the real shift in the consultation and the accompanying AI practices. Traditional vulnerability management assumes discovery is the bottleneck: teams scan, rank issues, wait for a change window and patch on a periodic cadence. The regulators’ AI thinking starts from a different premise. If advanced AI cuts the time, cost and skill needed to find and exploit weaknesses, discovery gets cheaper for attackers and the firm’s bottleneck becomes execution speed.
That is why the supporting effective practices on AI push continuous exposure management. Firms are asked to map the full exposure graph around critical services: AI systems, open-weight models, APIs and model endpoints, cloud services, connectors, hardware and software, human and non-human identities, third-party and nth-party dependencies, and infrastructure such as data centres, telecommunications, power, water and open-source components. The point is not a prettier asset register. It is to understand which attack paths matter most to important business services, where internet exposure or supplier concentration sits, and which legacy systems will slow containment.
The regulators’ concern is grounded in a threat environment that appears to be worsening, even if causation is not settled. The Bank of England says October 2026 recorded the second-highest monthly number of cyber incidents reported to the FCA in five years, 52% above the previous 18-month peak in May 2025. In a recent joint Bank/FCA survey, 88% of respondents said frontier-AI cyber capabilities were likely to have a moderate or high impact on their risk profile over the next two years, and 64% said vulnerability management and patching had seen the biggest increase in management attention because of frontier AI. Those figures measure expectations and incident volumes, not proof that AI caused the increase. Even so, they help explain why supervisors are focusing on response tempo.
A useful board-level scorecard falls out of that logic. Can the firm quickly identify affected assets and dependencies? Can it create a safe change window without breaking payments, trading or claims handling? Can it patch or isolate the problem, verify that the fix really reduced risk, revoke compromised access and restore service safely? That is a more demanding test than counting tickets closed or alerts generated.
Evidence, not alert volume
The consultation’s governance model is conventional in form and sharper in consequence. Firms are asked to make boards and senior managers accountable for cyber risk, designate a responsible senior manager, keep current inventories and maps of critical assets and dependencies, protect systems and data with proportionate controls, monitor for attacks, maintain tested response and recovery plans, regularly test defences and learn from incidents and exercises. The NCSC Cyber Assessment Framework is supposed to inform those assessments, and the whole approach is meant to scale with a firm’s size, complexity and risk.
What changes in practice is the kind of evidence boards may need to ask for. The AI paper warns that defensive tools can help identify and remediate exposures, but their outputs are non-deterministic and require validation and human oversight. It introduces “harness engineering” — the workflows, controls, data, tool access, validation steps, integrations and operating environment around an AI system that make its output usable and accountable. That is a useful corrective to the industry’s habit of evaluating tools by how much they find rather than by what they actually close.
For cyber teams, that means proving remediation, not merely detection. Did the vulnerability get disclosed to the right owner? Was the exploit path broken in production? Was the risky model state rolled back? Were unsafe memory or embeddings redacted? Could the team pause or kill an AI-driven process with human authorisation? Did immutable backups allow a clean restore? These are operational questions, but they also redistribute power inside firms. Procurement and vendor-management teams will need better answers from suppliers about assets, logging, software bills of materials, model provenance, recovery steps and cryptographic dependencies. Security vendors, in turn, may face more pressure to demonstrate validation and recovery support rather than just finding more issues.
This is also where the public-interest dimension becomes clear. In finance, a cyber problem at a cloud provider, payment rail, exchange or widely used open-source component can hit many firms at once. The proposed guidance therefore reaches beyond the compliance department. Customers may experience more intrusive evidence requests from banks and insurers, while critical suppliers may see tighter contract terms and deeper resilience due diligence.
What firms can do before the consultation closes
The formal status matters. The consultation runs until February 26, 2027, and the FCA expects final guidance in the first half of 2027. It is aimed mainly at UK banks, building societies, designated investment firms, Solvency II insurers and reinsurers, the Society of Lloyd’s and managing agents, enhanced-scope SM&CR firms, UK recognised investment exchanges, e-money and payment institutions, registered account-information service providers, consolidated-tape providers and, from October 25, 2027, qualifying cryptoasset firms. It does not formally apply to every FCA-regulated firm, and it does not create new binding rules.
Even so, firms do not need to wait for a final text to see the direction of travel. The most concrete near-term work is unglamorous: update asset and dependency maps; identify the business services that cannot tolerate delayed remediation; rehearse isolation, rollback and restore steps; decide who can authorise emergency changes; and collect evidence that fixes worked. For firms using AI in security operations, the regulators’ message is not “automate everything.” It is “use AI where it helps, but keep humans responsible for key decisions and validate the output.”
The quantum material, released alongside the consultation, points to the same habit of early inventory and long-horizon execution. Regulators are not saying quantum machines are breaking financial encryption today; the document explicitly says no cryptographically relevant quantum computer is known to exist. The warning is about migration lead times and harvest-now-decrypt-later risk. Firms are encouraged to build a cryptographic inventory, plan a prioritised migration, improve crypto-agility, test changes, coordinate with third parties and consider hybrid use of traditional and post-quantum algorithms during the transition. Accountability does not disappear when the cryptography sits inside a supplier’s product.
What remains unresolved is how supervisors will judge speed and sufficiency in practice. The materials do not set a single exposure-management metric, a maximum remediation time, an AI assurance threshold or a deadline for post-quantum migration. They do not say how much telemetry or rollback evidence vendors must provide, or how the final guidance will land for smaller firms and cross-border branches. That uncertainty is precisely why GC26/9 matters now. The consultation is not asking firms to predict the next tool trend. It is asking whether their cyber resilience can be demonstrated as an operating capability, with named accountability and tested recovery, before an attacker or an outage forces the proof.




By
By
By
By
By








