Artificial intelligence

Since the start of the year, authorities' attention has concentrated on four key AI risk watch points:

• Frontier AI capabilities that could accelerate cyber attacks and operational disruption across shared financial sector infrastructure
• AI-enabled fraud and impersonation that can make scams more persuasive, personalized and scalable
• Growing concentration and opacity in third-party AI, cloud and software supply chains
• AI-driven market, financial stability and infrastructure financing risks, including synchronized trading or redemption behaviour, stretched AI-sector valuations and opaque debt channels linked to data centers, chips and related infrastructure

Across these areas, the common concern is that AI may increase the speed, scale, interconnectedness and opacity of risk transmission, making threats harder for firms and supervisors to detect, attribute, contain and recover from before they affect consumers, markets or critical financial infrastructure.

Anthropic's April 7 release of Claude Mythos Preview was followed by a concentrated series of financial-sector warnings in the second quarter of 2026. The model was reported to have autonomously identified thousands of previously unknown vulnerabilities in widely used operating systems and web browsers, while a U.K. AI Security Institute assessment found that it was the first model to autonomously sequence a complete 32-step corporate-network attack, completing all steps in three of 10 attempts. Anthropic initially limited access to about 50 Project Glasswing partners focused on patching identified vulnerabilities and sharing non-sensitive findings. Authorities broadly warned that such capabilities could make sophisticated attacks faster, cheaper and accessible to less-skilled actors, narrowing the interval between vulnerability discovery, exploitation and remediation and increasing the risk of simultaneous failures across common software, cloud, payments and data infrastructure.

The International Monetary Fund said correlated failures could disrupt financial intermediation and payments and produce funding strains, solvency concerns, confidence effects and fire-sale dynamics. In Australia, the Australian Prudential Regulation Authority said information security practices were struggling to keep pace and flagged prompt injection, data leakage, insecure integrations, exploit injection, autonomous-agent misuse, incomplete security testing and patching timelines that did not match the accelerated threat; the Australian Securities and Investments Commission warned that isolated weaknesses could have a systemwide domino effect. The Securities and Exchange Board of India cited the potential for cascading effects across the interconnected securities market and established the cyber-suraksha.ai task force, while the Financial Conduct Authority, Bank of England and HM Treasury said current frontier models already exceed what a skilled practitioner could achieve at greater speed, larger scale and lower cost. Japan's Financial Services Agency and the Bank of Japan called for short-term preparations for the rapid discovery of large numbers of vulnerabilities and the resulting volume of patches, while South Korea's Financial Services Commission announced arrangements to support AI-driven cyber defense.

Across these actions, authorities reinforced expectations that boards and senior management treat frontier AI as a management and operational-resilience issue, accelerate vulnerability triage and patching, strengthen access, network and data controls, monitor third-party and open-source dependencies and maintain tested containment, incident response, recovery and business continuity arrangements.

During the first half of 2026, authorities increasingly described AI as an accelerant of fraud, allowing scammers to manufacture credible identities and endorsements, personalize approaches and maintain convincing interactions at greater speed and scale. Their warnings also show schemes moving beyond public advertisements and websites into closed messaging groups and private conversations, where AI-generated content and simulated social activity can build trust while making perpetrators harder to identify and investigate.

The Austrian Financial Market Authority identified deepfake celebrity videos, WhatsApp investment groups and AI-powered chatbots as emerging investment-fraud techniques in 2025. It described groups that promote trading tips and exclusive opportunities, use chatbots to simulate an active investment community, and then approach victims privately to direct them to fraudulent platforms or apps and overseas accounts or wallets. The authority received 843 fraud reports involving a record 19.6 million euros in losses, up from about 15.5 million euros the previous year, and issued 97 warnings about unauthorized providers, most linked to trading platforms. Separate warnings from the Bank of Italy and the Central Bank of San Marino illustrated the growing use of deepfakes to impersonate senior central bank officials: fabricated images, videos and articles presented the officials in television or other media settings and associated them with investment platforms or solicitations, prompting both institutions to file complaints with the relevant authorities.

Canada's Financial Industry Forum on Artificial Intelligence Phase II report broadened the concern to voice cloning, synthetic identities and personalized scams, noting that effective deepfakes can be produced from limited information obtained through social media and that call centers and help desks can be targeted through convincing impersonation of customers or employees. The report cited an industry survey in which 91% of financial institutions said they were reconsidering voice-verification systems because of AI voice cloning, and said personalized fraud using consumer data and behavioral patterns is becoming harder for consumers to distinguish from legitimate communications; it also cited an estimate that 90% to 95% of fraud in Canada goes unreported. In May, the U.S. House Financial Services Committee said AI could be weaponized to conduct scams at unprecedented scale and speed and advanced legislation addressing scams targeting older adults, federal assessment of AI-related fraud risks and access to advanced fraud-detection tools for smaller banks and credit unions.

Supervisory attention is increasingly focused on the systemwide effects of common AI and cloud dependencies, rather than solely on the resilience of individual vendors. The concern is that the same provider, model, platform or embedded software component may support critical functions across multiple institutions, while subcontracting chains and offshore hosting leave firms and supervisors with incomplete visibility into who supplies, trains, updates, secures or could replace the service.

In June, a European Central Bank Executive Board member said banks increasingly rely on external providers for critical functions that are difficult or impossible to replace, exposing them to cascading effects from supply-chain incidents even when they are not directly targeted. He noted that more than 85% of banks under European banking supervision already use AI and pointed to the Digital Operational Resilience Act's enhanced oversight of critical third-party providers, including cloud service providers. France's Financial Markets Authority also placed third-party information and communications technology risk within its June supervisory focus under the act, highlighting the mapping of systems and suppliers and announcing plans to survey firms in July on measures taken or planned in response to AI-related risks.

The Australian Prudential Regulation Authority's April review provided a firm-level view of the same issue: Some banks, insurers and superannuation trustees were heavily dependent on a single provider for multiple AI applications, few had demonstrated robust contingency plans or tested exit and substitution strategies, and contracts often lacked provisions covering audit rights, model changes, incident notifications and changes to data handling. The authority also found that AI embedded in software platforms and developer tools could obscure the underlying foundation models, training data and fourth-party providers, limiting institutions' ability to assess model performance, resilience and security.

A March report from the Canadian Financial Industry Forum on Artificial Intelligence described AI dependencies as extending across data, models, open-source software and compute and cloud infrastructure. It flagged limited visibility into fourth- and fifth-party relationships, uncertainty about how models were trained and what data they used, and the placement of critical providers and data outside the financial regulatory perimeter or in foreign jurisdictions. Likewise, a report by the South African Reserve Bank, through its Prudential Authority, said that only a handful of companies have the technology and resources to train and operate AI models, making outsourcing increasingly widespread and creating single-point-of-failure and sensitive-data exposure.

Recent research and official statements have sharpened two distinct financial stability concerns: how autonomous AI systems may behave during periods of stress and how the expansion of AI infrastructure is being valued and financed.

In May, researchers writing in a European Central Bank publication used a simulated mutual fund redemption problem to show that different AI architectures could produce materially different forms of instability. Q-learning agents converged on excessive redemptions under default risk even when economic fundamentals were strong, while large language model agents were less prone to such runs but formed differing beliefs about other investors, making outcomes less predictable. The authors said the same dynamics could apply to bank runs, currency attacks and stablecoin runs.

In June, the South African Reserve Bank's Financial Stability Review identified a separate AI-related asset valuation channel, saying the expansion of AI capabilities had supported sustained growth in technology-related share prices and raised concerns about stretched valuations and the potential for a disorderly correction. Scrutiny of AI infrastructure financing also intensified in the United States. In January, Democratic members of the Senate Committee on Banking, Housing and Urban Affairs asked the Financial Stability Oversight Council to investigate risks associated with more than USD 1 trillion in projected debt financing for AI infrastructure, citing the growing use of private credit, securitizations and off-balance-sheet structures to fund data centers and related investments. They warned that these arrangements could obscure leverage and expose banks, insurers, private credit funds, real estate investment trusts, pensions and retail investors if AI companies fail to generate enough revenue to service their debt.