Commentary | Euro Security EMEA

A convincing voice, a familiar face and an apparently legitimate payment instruction may no longer be sufficient evidence of identity. Generative artificial intelligence is exposing a fundamental weakness in financial security: institutions have become increasingly sophisticated at protecting digital transactions while remaining vulnerable to the manipulation of the people authorised to approve them. OpenAI CEO Sam Altman’s warning about a potential fraud crisis has attracted considerable attention. Yet the more uncomfortable question is not whether AI could undermine banking security. It is whether financial institutions, regulators and technology providers are responding quickly enough to a threat that has already produced substantial losses.

The problem is not a banking collapse. It is counterfeit authority.

When Sam Altman addressed a Federal Reserve conference in Washington on 22 July 2025, he expressed concern that some financial institutions still accepted voice recognition as sufficient authentication for sensitive transactions. Advances in artificial intelligence, he argued, had made such arrangements dangerously vulnerable to impersonation. His warning was subsequently transformed into dramatic claims that the global banking system had effectively collapsed and that mass withdrawals from customer accounts were imminent. Neither conclusion is supported by the available evidence. Modern banking security has not ceased to function, and strong customer authentication, transaction monitoring and fraud detection systems continue to prevent substantial amounts of financial crime. What has changed, however, is the credibility of the signals on which people often rely to establish identity. A familiar voice can be synthesised, a senior executive digitally impersonated and a video conference manipulated to create the appearance of legitimate authority.

The implications extend beyond traditional account takeover. Criminals increasingly have opportunities to manipulate authorised individuals into performing actions that appear legitimate to the systems processing them. A transaction may therefore be technically authenticated and still be fraudulent. This distinction is becoming one of the defining security challenges of the AI era. Financial institutions cannot address it simply by improving facial recognition or adding another layer of biometric verification. They must reconsider the assumptions underpinning the entire authorisation process, particularly where visual or acoustic familiarity is still treated as convincing evidence of identity. The fundamental weakness is not necessarily a failure of encryption, authentication software or banking infrastructure. It is the continued reliance on human recognition and apparent authority in situations where those signals can be convincingly fabricated.

Hong Kong demonstrated how expensive synthetic trust can become

The danger is no longer hypothetical. In January 2024, an employee at the Hong Kong office of international engineering consultancy Arup was deceived into transferring approximately HK$200 million, equivalent to around US$25 million, following a fraudulent video conference. According to information subsequently released by the Hong Kong authorities, criminals impersonated the company’s UK-based chief financial officer and other colleagues using AI-manipulated video and audio material. The authorities later concluded that the meeting had been pre-recorded rather than genuinely interactive. The employee authorised transfers to five bank accounts, and Arup subsequently confirmed that it had been the victim of the fraud. The case illustrates a particularly uncomfortable reality for corporate security: the criminals did not necessarily need to compromise the company’s banking infrastructure. Instead, they created a credible representation of corporate authority and persuaded an authorised employee to act on it.

Organisations can protect networks, enforce access restrictions and require multifactor authentication, yet these safeguards may not prevent financial losses when the person who legitimately controls a transaction has been deceived. Adding a video call to an existing approval process does not necessarily make that process more secure; it may simply provide another opportunity for impersonation. For multinational organisations, the implications are especially serious. Financial approval chains frequently cross national boundaries, involve remote executives and depend on digital communication between offices operating in different time zones. A fraudulent instruction can therefore exploit not only technological weaknesses but also organisational complexity. In such environments, trust must be established through independently verifiable procedures rather than the apparent familiarity of a voice, face or communication channel. The Arup case should therefore be understood not merely as an example of sophisticated digital fraud, but as a warning about the fragility of approval processes that depend too heavily on perceived personal authority.

International fraud figures expose the limits of authentication

The scale of financial fraud is substantial, although available statistics must not be misrepresented as evidence that all losses result from artificial intelligence. In the United States, the FBI’s 2025 Internet Crime Report, published in April 2026, recorded more than 22,000 complaints involving AI-related information, with reported losses exceeding US$893 million. These figures include different forms of AI-assisted fraud and do not represent losses from voice cloning or deepfakes alone. In Europe, a joint report by the European Banking Authority and the European Central Bank recorded €4.2 billion in payment fraud across the European Economic Area in 2024, compared with €3.5 billion in 2023. The institutions confirmed that strong customer authentication remained effective, particularly against traditional card payment fraud. However, they also identified growing risks from schemes in which legitimate customers are manipulated into authorising fraudulent transactions. For credit transfer fraud, payment service users bore approximately 85% of the reported losses.

Singapore provides another revealing perspective. According to the Singapore Police Force, financial losses from scams reached approximately S$913 million in 2025, although this represented a decline from the previous year. In 81.8% of scam cases, victims had initiated the transfers themselves after being deceived. Again, these are broader scam statistics, not measurements of AI-generated impersonation. Nevertheless, the findings from three different jurisdictions point towards a common vulnerability: criminals can exploit the gap between technical transaction authorisation and the authenticity of the circumstances that produced it. The next major banking fraud may not begin with a stolen password but with a properly authenticated customer making a decision based on fabricated reality. This is precisely why financial institutions must look beyond the technical validity of a transaction and examine the circumstances under which it was initiated.

Banks cannot outsource the consequences of deception to customers

The distinction between unauthorised transactions and authorised payments induced by fraud is central to the international debate about financial responsibility. Historically, banks have concentrated heavily on determining whether the person executing a transaction was authorised to do so. But what happens when that person has been manipulated by an apparently credible executive, bank employee or family member? If the correct customer initiates the payment, the correct authentication method is used and the transaction is processed according to established procedures, the technical controls may perform exactly as designed. The financial outcome can still be devastating. This creates fundamental questions of accountability: Should a customer bear the consequences of increasingly sophisticated deception? To what extent should banks be expected to detect unusual transaction patterns, and what responsibilities should apply to receiving institutions that accept fraudulent funds?

Different jurisdictions are beginning to approach these questions differently. The United Kingdom introduced mandatory reimbursement requirements for eligible victims of Authorised Push Payment (APP) fraud in October 2024. Under the framework established by the Payment Systems Regulator, qualifying consumers, microenterprises and charities can receive reimbursement for covered domestic bank-transfer scams, generally subject to a maximum of £85,000 and specified exceptions. The rules place stronger incentives on payment service providers, including receiving institutions, to prevent fraud rather than relying exclusively on customers to recognise deception. The European Union has pursued stronger authentication and payment security requirements, while its financial supervisory authorities continue to examine the limitations of existing controls against manipulation-based fraud. Across the Middle East, Africa and Asia, financial institutions operate under different national regulations, supervisory expectations and customer-protection arrangements. A single international liability framework does not exist, although the underlying security problem is global. Digital communication, AI-generated identities and cross-border payment networks do not respect the boundaries of individual regulatory systems. Where a transaction presents clear indicators of manipulation or abnormal behaviour, a successful authentication check should not automatically end the institution’s responsibility to investigate. Customer vigilance matters, but it cannot remain the industry’s only defence.

More biometrics will not necessarily deliver more security

One tempting response to AI impersonation is to demand stronger biometric identification. That approach requires caution because biometric authentication, while potentially effective within secure technical environments, is not inherently equivalent to verifying the cryptographic identity of a device or the legitimacy of a transaction. The distinction becomes critical when attackers can generate convincing synthetic representations of real people. A voice recording cannot be treated as a secret, while facial characteristics are not equivalent to a password that can be replaced after compromise. Even sophisticated liveness-detection systems must be evaluated against evolving manipulation techniques. No single measure should be assumed to provide permanent protection against every form of synthetic media. Increasing the number of biometric checks without reconsidering the underlying authentication architecture may therefore create additional complexity without adequately addressing the central threat.

Financial institutions need security architectures that remain reliable even when individual identity signals are compromised. These should combine phishing-resistant authentication, transaction-specific confirmation, independent approval channels and additional verification for unusual changes to accounts or payment arrangements. For corporate payments, segregation of duties and independently verified approval procedures remain essential. An instruction supposedly issued by a chief executive should not become legitimate merely because it arrives through a familiar communication channel. Technology can help identify suspicious behaviour, but the broader objective must be to remove single points of failure from critical financial decisions. The industry must stop confusing the ability to recognise someone with proof that the person is genuinely present, authorised and acting under legitimate circumstances. Security must be designed to remain effective even when a convincing impersonation succeeds.

Responsibility extends beyond the financial industry

The debate should not end with financial institutions. Developers of generative AI systems are creating technologies with substantial legitimate applications in communication, accessibility, entertainment and customer service. The same capabilities can support identity impersonation, fraudulent instructions and scalable social engineering. It would be simplistic to argue that AI developers alone are responsible for every criminal application of their technologies, since criminal conduct remains the responsibility of those carrying it out. Yet it would be equally inadequate to treat misuse as an entirely external problem. Technology providers should be expected to investigate abuse patterns, strengthen relevant safeguards and cooperate with financial institutions, regulators and law enforcement where appropriate.

Digital platforms and telecommunications providers also have a role in disrupting fraudulent communications and limiting the misuse of their services. The challenge is to establish practical responsibilities across the financial and digital ecosystem, from the creation and distribution of synthetic content to the authorisation and movement of funds. This requires international cooperation because fraudulent operations can involve several jurisdictions, service providers and payment networks within a single incident. The widespread availability of generative AI also means that banks cannot base their security assumptions on the hope that the most convincing impersonation technologies will remain inaccessible to criminals. Financial institutions, technology developers and regulators must instead work from the assumption that convincing synthetic identities will increasingly form part of the threat environment. Effective protection will depend on their ability to combine technical safeguards, fraud intelligence and clearly defined institutional responsibilities.

The real test is whether security works after deception succeeds

The financial industry has spent decades investing in the confidentiality and integrity of digital transactions. Generative AI is now forcing institutions to confront a different dimension of security: the integrity of the information on which authorised people make decisions. This shift demands more than technical upgrades. It requires banks to examine how identity is established, how high-risk actions are approved, how suspicious payments are challenged and how quickly fraudulent transfers can be interrupted or recovered. It also requires a change in corporate culture. An employee who requests independent confirmation of an urgent instruction should not be treated as obstructive, just as a customer who hesitates before authorising an unusual transaction should not be pressured into faster compliance. Security procedures should be designed to withstand social pressure rather than depend on an individual’s ability to resist it.

Regulators, meanwhile, should evaluate fraud prevention not only by the strength of authentication technologies but also by the effectiveness of transaction monitoring, customer protection and incident response. Sam Altman’s warning at the Federal Reserve did not establish that the global banking system was on the verge of collapse. It exposed something more specific and arguably more important: a growing mismatch between the sophistication of digital impersonation and the assumptions embedded in some traditional identification practices. The response should be proportionate to the evidence, but not complacent. A convincing face is no longer proof of identity, a familiar voice is no longer proof of authority, and a correctly authenticated payment is not necessarily a legitimate one. The financial institutions that understand this distinction will be better positioned to defend their customers and operations against the next generation of fraud. Those that continue to place the burden of recognising increasingly convincing deception on individual customers and employees risk discovering that their systems can operate perfectly while their security strategy fails.

Editorial sources: This commentary draws on the official Federal Reserve transcript of Sam Altman’s remarks on 22 July 2025; the Hong Kong Government’s account of deepfake-enabled corporate fraud; the FBI’s 2025 Internet Crime Report; the joint EBA–ECB payment fraud report published in December 2025; the Singapore Police Force’s 2025 scam statistics; and the UK Payment Systems Regulator’s rules on reimbursement for authorised push payment fraud.