Five lessons in payment security that AI hasn’t made obsolete

Firms should emphasise extending control areas that already underpin trust in payments to cover AI systems

Summarise
阅读简体中文版 (beta)
    • The controls that already secure card data and payment infrastructure, applied with the same rigour to AI, close most of the gap that matters.
    • The controls that already secure card data and payment infrastructure, applied with the same rigour to AI, close most of the gap that matters. IMAGE: PIXABAY
    Published Wed, Oct 28, 2026 · 07:00 AM

    SINGAPORE’S banks and fintech companies are using artificial intelligence in live payment operations, putting the country ahead of most of its regional peers that are still running pilots, a 2026 survey by fintech company Finastra showed.

    That lead signals what is coming for the rest of the region, as AI takes on a growing share of decision-making, processing and fraud detection across payment systems.

    Where the real risk sits

    Most fraud controls in payments today are built to answer this question: Is the person who they claim to be?

    As AI takes on a larger role in initiating and approving payments, that question becomes harder to answer, because the systems flagging anomalies were designed around human behaviour, not automated technology operating with already-granted permissions.

    The more consequential shift currently under way concerns process, not detection.

    An AI system can operate exactly as it was configured to, and still produce a financially damaging outcome, because the question being asked was whether the process was authorised, not whether it should have been permitted in the first place.

    That is a different failure mode from the one that payment security has spent two decades defending against, and one that most stress-testing exercises are not yet designed to catch.

    Sumsub’s 2026 Asia-Pacific fraud report found that synthetic identity fraud went up 142 per cent year on year, and automated attacks rose 388 per cent.

    Separately, Mastercard’s research on Apac small and medium-sized enterprises found that 47 per cent rank fraud and security protection as the AI capability they most want from their financial providers.

    Yet, only 39 per cent of these SMEs use cybersecurity tools, despite 70 per cent calling it a high priority.

    The appetite for AI-powered defence is there; the controls to match it, in many organisations, are not.

    Singapore’s regulators are already onto this.

    The Monetary Authority of Singapore’s proposed guidelines on AI risk management is pushing accountability for AI risk to board level, and specifically calling out prompt injection and data leakage as risks financial institutions must actively manage, not simply monitor.

    What organisations should actually do

    The Payment Card Industry (PCI) Security Standards Council’s approach – set out in the September information supplement, Security Considerations for AI Systems – is deliberate.

    Rather than treating AI as needing an entirely new rule book, the emphasis is on extending the same control areas that already underpin trust in payments, including access, data protection, logging and incident response, to cover AI systems specifically.

    Assessors and merchants already understand this framework; applying it consistently to AI use closes the gap faster.

    In practice, the most common failure has little to do with inventive new attack types. It comes down to AI systems being handed more access and autonomy than the controls around them were built to handle.

    The five lessons

    First, treat AI within its existing controls, not around them.

    Wherever AI touches cardholder data or payment infrastructure, the fundamentals should apply in full: least-privilege access, data protection at rest and in transit, and logging that is detailed enough to reconstruct what a system did and why.

    AI systems are not exempt from these.

    Second, keep a named human accountable for the outcome. AI can inform or execute a task, but responsibility for the result has to sit with a person, not a system.

    Clear sign-off points, and incident response plans that account for an AI system behaving unexpectedly or being manipulated, close that accountability gap.

    “The AI did it” cannot become an acceptable answer during an audit or a breach investigation.

    Third, validate continuously, rather than only at launch.

    AI systems drift and behave inconsistently as they take in new data. A control that passed testing six months ago is not necessarily still doing what it was built to do, so validation has to be treated as an ongoing operational necessity.

    Fourth, design for containment and define the limits of AI use. Give AI systems only the access and credentials a task requires, ensure that access can be revoked quickly, and build in a straightforward way to disable the system if something goes wrong.

    A basic guiding question: What happens when this fails?

    At minimum, an AI system should not have access to unencrypted secrets, should not make decisions that no human can be held accountable for, nor have access to systems not required for its operation.

    Fifth, harden against manipulation. Payment-facing AI is now a target for prompt injection and similar techniques that extract data or trigger fraudulent actions.

    The inputs an AI system accepts and the outputs it can act on both need the same scrutiny as any other exposed interface, because AI systems are often trusted with more implicit authority than comparable traditional software.

    Organisations should resist the temptation to treat AI as a special case requiring bespoke governance, and instead ask an older, simpler question: Does this system meet the same standard of care that everything else in the payment environment is already held to?

    For most organisations in Singapore and across the region, the honest answer is: no, not yet.

    Closing that gap requires attention to both the security of AI systems within payment environments and the security of traditional systems against AI-enabled attacks.

    The controls that already secure card data and payment infrastructure, applied with the same rigour to AI, close most of the gap that matters.

    This is ultimately a test of whether organisations are willing to apply the rules they already have to a part of their environment that is moving faster than their oversight of it.

    The writer is regional vice-president for the Asia-Pacific; Europe, Middle East, and Africa; and Latin America and the Caribbean at the PCI Security Standards Council, based in Singapore