When Visa — processing $15 trillion in payments annually — admits that its own AI-powered defences have vulnerabilities that rival AI systems can exploit, the rest of the financial services sector should stop scrolling. This isn't a technical curiosity. It's a direct threat to any UK firm handling regulated client data, customer funds, or sensitive commercial information. The FCA Consumer Duty (PS22/9) and PRA SS1/23 both demand that firms understand and manage emerging risks. Autonomous cyberattacks powered by AI aren't emerging anymore. They're here. And Visa's open-sourcing of part of its defence stack — effectively saying 'we need the industry to help us fix this' — is a clear signal that the vendor-lock-in approach to AI security has hit a wall.

What we're seeing is a pattern: major financial institutions are discovering that AI-first security strategies contain the same fundamental flaw as the threats they're meant to stop. An AI model trained to detect anomalies in payment flows can itself be probed, tested, and potentially fooled by another AI model. It's not a bug in Visa's system specifically. It's a category error in how the industry has approached AI security. Microsoft Copilot, Harvey, Luminance and other large language models are now standard in financial services and legal firms. Each introduces new attack surfaces. Each requires governance that most mid-market firms simply don't have. The shift toward autonomous cyberattacks — and the quantum-computing threat Visa mentions — means that the old perimeter-defence model is dead. But the AI-security-as-solution model is dying too. What's left is messy, distributed, and requires transparency that most vendors hate to provide.

Here's Trovix's honest take: AI can detect patterns human analysts miss, but AI-to-AI combat is still in its infancy, and betting the firm on proprietary black-box models is reckless. The answer isn't less AI. It's structured AI, explainable AI, and — crucially — human-centred governance. The EU AI Act (Article 8 high-risk classification for financial services and insurance) and ICO UK GDPR guidance both push toward this. Firms using tools like Luminance for document review benefit because the tool flags its reasoning. Tools like Harvey that are purely generative without transparency create liability. What we'd argue at Trovix is simpler: your AI systems need to be auditable. You need to know why they flagged something. You need to know when they've been tested against adversarial input. Trovix Audit exists precisely because we saw firms deploying AI without any governance layer. Visa's move tells us we were right.

For a mid-market financial services firm, insurer, or law firm: stop treating AI security as something your vendor handles. Map your AI dependencies — where are you running language models, what data are they touching, who has access? Run a vulnerability review not of your networks, but of your AI systems specifically. Get independent assessment of any AI tool that touches regulated data. The SRA Code of Conduct for law firms now includes specific obligations around technology competence and bias. Your PRA or FCA regulator will expect evidence of this. Visa's confession is your early warning. Act now, while you can still control the narrative.

Source: Thomson Reuters