Deep Dive AI governance & cyber resilience
It was never about the model. It’s about the harness.
Two supervisory documents landed two days apart at the turn of the month. Together, they make an uncomfortable case: the question is not whether the model can find a vulnerability. It is whether the organisation around it can absorb what it finds.
FCA publishes its first multi-firm review of frontier AI in cyber resilience.
FSB Chair writes to G20 finance ministers on frontier AI and correlated cyber disruption.
A diagnostic of whether existing resilience arrangements can cope with AI-accelerated discovery.
For two years, much of the risk conversation has sat in the model: can it scan safely, be fooled, miss something, or create new exposure? The FCA’s review shifts the unit of assessment. Vulnerability discovery is accelerating faster than firms can respond, and whether that helps or harms a firm depends on the governance, tooling, controls and human oversight around the model—the harness.
The FSB makes the same point at a different level. Frontier AI may alter the speed, scale and economics of cyber risk. The resulting question is not only whether one firm can cope with more discoveries; it is whether several firms can cope when they share a cloud region, a security provider, a model provider or another critical dependency.
The queue is the new attack surface.
The FCA’s plainest finding is arithmetic, not technical. AI-assisted tools can surface genuine vulnerabilities faster than remediation teams can work through them. Even after triage, the remaining volume can place serious pressure on engineering capacity, patch testing, change management and the evidence needed to close an issue.
Most vulnerability programmes assume a discovery rate that broadly tracks headcount and established tooling. That assumption no longer holds. Scaling the scanner without changing the authority and capacity behind the queue simply widens the gap.
The control response is re-sequencing authority: someone with standing to decide, in real time and on a documented basis, which findings jump the queue and which wait.
QuestionIf AI-assisted scanning tripled your open-vulnerability count tomorrow morning, who could re-sequence the queue—and would that decision survive an audit?
The model is not the control. The harness is.
The FCA reports that the value and risk of frontier AI are determined less by the underlying model than by the operating environment around it: governance, specialist tooling, validation, access boundaries, human approval and the information available to the people using it.
That reframes two years of due diligence. The assurance question cannot stop at the provider’s own documentation or benchmark results. It must map the practical guardrails between the tool and production systems, including permissions, escalation routes, the quality of system and asset context, and the people who can intervene.
Audit the harness, not the vendor. Produce a control map that stands independently of supplier assurance.
QuestionCould your team produce a control map of its frontier-AI harness this week, including governance, access boundaries and human checkpoints?
Asset hygiene is now a cyber-resilience finding.
Faster discovery exposes weaknesses in asset mapping, dependency management and remediation processes that can remain hidden when the volume is manageable. Under AI-assisted discovery, a stale technology-debt entry can become a live control problem very quickly.
The relevant control environment is not a set of supporting documents. It is the working connection between the asset register, the important-business-service map, ownership, vulnerability triage, emergency change and evidence of closure.
Treat asset and dependency inventories as board-reportable controls in their own right, not as supporting material for something else.
QuestionWould this quarter’s asset register hold up if a frontier-AI tool began finding, accurately, everything it currently misses?
Third parties will not automatically keep pace.
The FCA explicitly points to supplier preparedness, cloud dependencies, software supply-chain visibility and shared infrastructure. A firm can improve its own harness, but it cannot assume that a critical supplier’s patch cadence, validation process or escalation path will match an AI-accelerated threat tempo.
The exposure does not disappear when it crosses the boundary. It changes owner, evidence route and recovery option—and often becomes harder to test.
Build and rehearse an explicit escalation path with critical suppliers for AI-accelerated findings. A contract clause is not a response process.
QuestionHave you asked your most critical suppliers whether their remediation service level still holds against an AI-accelerated threat?
Prepare for more severe scenarios involving simultaneous disruption across multiple firms or shared technology dependencies.
The systemic layer: correlated failure across firms.
The FCA’s review is necessarily about a firm’s own cyber-resilience posture. The FSB’s letter asks what happens when many firms share the same AI-accelerated exposure. That creates a scenario most resilience programmes do not yet test: not simply “what if we go down?”, but “what if we and several peers go down together because we share a vendor, model, cloud region or security tool?”
The board communications challenge matters as much as the technology. “We did not cause it” is not an adequate explanation if the firm cannot show that it understood its dependencies, knew how to intervene and could demonstrate safe recovery.
Extend severe-but-plausible scenarios to include correlated, shared-dependency disruption—and rehearse the board communication for one.
QuestionDoes the scenario library include a shared-vendor or shared-model failure that affects several peer institutions at once, not merely a single-firm outage?
For the board
Three questions worth asking this quarter
- 01
Who owns re-sequencing authority when AI-assisted scanning outpaces the existing remediation queue—and is that decision documented?
- 02
Can the team produce a harness control map—governance, access boundaries and human checkpoints—independent of vendor assurances?
- 03
Does the severe-but-plausible scenario library include correlated, multi-firm disruption through a shared vendor, model or cloud region?
Sources