FCA frontier AI: cyber resilience evidence checklist.
An AI tool can produce a vulnerability finding in seconds. The difficult part is deciding whether the finding is valid, which service it threatens and who can safely fix it. Financial firms can use a small, traceable evidence pack to connect those decisions.
Regulatory snapshot
- Status: published multi-firm review, checked on 30 September 2026.
- Source: FCA, Frontier AI and cyber resilience.
- Audience: technology, cyber, risk and operational-resilience leaders, including smaller financial firms.
- Implementation deadline: none created by this publication.
What changed, and what did not
The new contribution is evidence from firms using or preparing to use advanced AI in cybersecurity. The review connects effective use with governance, human judgment and the surrounding operating environment. Its observations can inform a firm's own risk assessment; they are not a certification standard.
Existing operational-resilience duties remain separate. Under PS21/3, in-scope firms had to complete sufficient mapping and testing to remain within impact tolerances by 31 March 2025. That date has passed. An AI adoption project does not reset it.
Who should use this checklist?
Start with the legal entity and its activities. The FCA's operational-resilience framework covers specified firms, including banks, insurers and certain payment and investment firms; it does not apply identically to every technology supplier. Check the current FCA scope and operational-resilience requirements before recording a mandatory obligation.
For a UK banking compliance team, this is a useful joint exercise for cyber operations, service owners, risk and compliance. A supplier can use the same evidence structure to explain its controls to a client without representing itself as directly subject to every rule that binds the client.
A suggested evidence pack for one AI-assisted workflow
The following records are our recommendations. Select one workflow, such as finding weaknesses in an internal application, and test whether another reviewer can reconstruct the decision without interviewing the original operator.
| Control question | Suggested record | Suggested owner |
|---|---|---|
| What can the tool reach? | Approved system boundary, permitted actions, credential owner and expiry; separate discovery access from change permissions. | Security engineering |
| Can the finding be reproduced? | Finding identifier, affected asset, relevant tool configuration, validation result and reviewer decision. | Vulnerability management |
| Which service could be affected? | Link to the service map, dependency owner and impact rationale; mark uncertain links for review. | Business-service owner |
| Who can accept or fix the risk? | Named decision-maker, remediation ticket, due date, exception rationale and escalation trigger. | Risk and change owners |
| Can the firm recover from a bad change? | Test result, rollback decision, recovery evidence and approval before closure. | Operations |
A finding count alone is a weak completion measure. In a pilot, compare validated findings with the number that reach a documented fix or an authorised risk decision. Record the starting backlog so an increase in discovery is not mistaken for a sudden deterioration in security.
Work through a finding before expanding access
Consider a hypothetical scanner that flags an outdated library in a payments application. The analyst first confirms that the vulnerable version is actually deployed and reachable. The service owner then checks the dependency map. A proposed patch goes through the firm's change process, with a tested rollback available. The record closes only when the team has verified the result or an authorised owner has accepted the residual risk.
If validation cannot reproduce the issue, keep that conclusion and its evidence. If the dependency map is incomplete, assign a mapping task. Neither outcome justifies giving the scanner broader production permissions. This example is an implementation recommendation, not a reported FCA case or a claim about a Sia client.
Keep framework mappings honest
A group operating in several jurisdictions can reuse an evidence record while retaining separate legal assessments. A UK resilience record may be relevant to a DORA gap analysis, but it should not automatically close an EU obligation. Record the actual requirement, applicable entity and reviewer's reason for accepting the evidence in each mapping.
The NIST AI Risk Management Framework can provide an organising vocabulary for a voluntary AI assessment. Keep that reference separate from binding requirements and from internal policies that the firm has chosen to adopt.
How Sia RegAI can support the workflow
With the relevant sources configured, RegReview can help retain the publication, its status and subsequent changes. RegMatcher can support mapping applicable requirements to policies and evidence for review. Record this FCA publication as observations, so recommendations are not inadvertently converted into mandatory obligations.
The platform does not validate vulnerabilities, execute patches or decide acceptable cyber risk through this regulatory-mapping workflow. Technical specialists and accountable owners make those decisions. A useful demonstration starts with one source, one internal procedure and a sample evidence record whose limitations are understood.
What to watch next
Recheck the publication when the FCA issues a substantive update. Within the firm, repeat the assessment when tool permissions, suppliers or the affected service change. Keep the source version and the review date with the mapping; a record that was adequate for a read-only pilot may be insufficient for a tool authorised to make changes.
Primary sources
- FCA: Frontier AI and cyber resilience, 2 September 2026.
- FCA: PS21/3, Building operational resilience.
- FCA: operational-resilience scope and requirements.
This analysis is general information, not legal advice. The suggested records and workflow are Sia's recommendations. Confirm current rules, entity scope and technical controls with qualified owners.