On 2 September the FCA published something that deserves more attention than it has had: a multi-firm review concluding that frontier AI is now finding security vulnerabilities in financial services firms faster than those firms can fix them. The regulator calls this a "vulnerability bottleneck," and it is worth sitting with that phrase, because it names a problem most cyber strategies were not built to handle.
Discovery was never the constrained resource
The review's logic is straightforward once you see it. Frontier AI tools are very good at discovery. They can chain together several low-severity flaws that no individual test would flag, into an exploitation path that a traditional scan would miss entirely. That capability is genuinely useful. But discovery was never the constrained resource in cyber resilience — remediation was, and remediation still runs on people, change management processes and prioritisation judgement that have not got any faster just because the tool finding the problems has. The FCA's own words are blunt: "frontier AI is becoming a test of organisational resilience, not just a tool." Firms that read this review as validation of their AI security spend have misread it. It is a warning that the spend on discovery has outpaced the capacity to act on what is discovered.
A reclassification happening in real time
What makes the timing sharper is what happened two days earlier. On 31 August, Andrew Bailey — in his role as chair of the Financial Stability Board, not just governor of the Bank of England — wrote to G20 finance ministers naming frontier AI's effect on cyber risk as the "most immediate concern" to global financial stability, ahead of more familiar worries like stretched valuations and sovereign debt fragility. Put those two documents side by side and you get a genuine reclassification, happening in real time, at both the international and domestic level: AI has moved from a technology risk sitting in someone's IT risk register to a financial stability risk sitting in the same category as leverage and market concentration.
That is not a rhetorical shift. It changes who in an organisation should own the question and how often the board should be asking about it.
What a board should actually do with this
Not run out and buy more AI-enabled vulnerability scanning — most firms already have some version of that, and the review is explicit that model sophistication is not the differentiator. The differentiator is whether a firm can prove it remediates at the pace its tools now discover. That means named ownership for AI-enabled vulnerability triage, sitting somewhere specific rather than diffused across security, technology and third-party risk teams. It means mapping the assets and dependencies that sit behind critical services, so that when a chained vulnerability turns up, someone already knows what it touches. It means engaging critical suppliers on the same question, since a firm's remediation capacity is only as good as the weakest link in its supply chain. And it means testing that capacity under realistic volume, not just describing it in a policy document that has never been stress-tested against a busy week.
Better detection, without matching investment in the boring parts of remediation, does not reduce risk. It can increase it — by generating a backlog of known, unpatched weaknesses that everyone in the building can see on a dashboard and nobody has the capacity to close.
The question boards need to answer
The uncomfortable truth in this review is that better detection, without matching investment in the boring parts of remediation, does not reduce risk. It can increase it, by generating a backlog of known, unpatched weaknesses that everyone in the building can see on a dashboard and nobody has the capacity to close. If your board's honest answer to "how fast do we patch what our AI finds" is "we don't currently measure that," you have found your next agenda item.