Crédit Mutuel Arkéa reported that attackers exploited access through a service provider’s platform to transfer approximately EUR 100 million ($112.6 million) from the French banking group using its own account. The incident prompted the suspension of selected transactions, an investigation by French authorities and efforts to recover the funds. Elsewhere, Westpac cut the estimated cost of its St.George data warehouse migration by half after using AI to accelerate code conversion and testing, while National Australia Bank (NAB) reported measurable productivity gains across lending, software development and cybersecurity. Allica Bank strengthened validation controls for AI-assisted lending, while US and Indian authorities highlighted risks from delegated AI authority and attacks on financial infrastructure. Other developments included generative AI in fraud investigations at Mashreq and new frameworks for governing AI agents at Mizuho and in Hong Kong. Read more on the week’s key developments: 1. Crédit Mutuel Arkéa says attackers diverted $112.6 million through its own account French lender Crédit Mutuel Arkéa disclosed on 4 November that attackers had reached the banking group through a service provider’s platform and used the bank’s own account to transfer money to other banks. On 5 November, Arkéa put the amount stolen at approximately EUR 100 million ($112.6 million), according to Reuters. The Paris prosecutor’s cybercrime unit opened an investigation after receiving the bank’s complaint. Customers experienced delays after Arkéa suspended selected transactions to prevent further outflows, although balances at its member banks and online subsidiary Fortuneo remained intact. Almost all operations resumed on the morning of 4 November after the group and its subsidiaries checked their systems and strengthened security. Arkéa said it had blocked some of the funds and was pursuing recovery, leaving its eventual financial loss undetermined. The incident demonstrates how compromised third-party access can extend beyond data exposure to the movement of funds through a bank’s own infrastructure. The suspension of selected transactions also illustrates the operational disruption that can follow when a bank has to contain unauthorised payment activity while maintaining customer services. 2. Westpac halves St.George data migration cost to $24.5 million million using AI Westpac reported on 2 November that moving its St.George data warehouse would cost about AUD 35 million ($24.5 million), half the original AUD 70 million ($49.1 million) budget. Chief executive Anthony Miller told investors that engineers used AI to convert legacy code, test the results and document the changes. He said the team completed the migration more than two years ahead of schedule. Westpac had disclosed the expected time saving in September. Its November results quantify the project cost and identify when it should affect earnings. The bank moved the data to its cloud analytics platform and the business rules to Westpac Intelligence. It plans to retire the old warehouse in fiscal year 2027, when Miller expects it to start recognising the projected AUD 15 million ($10.5 million) in annual benefits. 3. NAB reports AI cuts of up to 90% in selected banking and technology tasks National Australia Bank (NAB) disclosed on 5 November that its AI tools were reducing processing times across several banking operations. Staff using AI recorded an average 20% reduction in business credit memorandum preparation time and a 75% reduction in the time spent identifying corporate and institutional banking customer insights. The bank also reported 50% less time for design-to-code delivery, 80% faster cyber incident triage and a 90% reduction in technology cyber assessment time. Chief executive Andrew Irvine said more than 40,000 employees would have access to Ruby, its AI assistant built on Anthropic technology. The results cover specific activities and user groups rather than bank-wide productivity. NAB's approximately 50 AI squads are developing applications for individual business processes, while agents for complaints handling, fraud detection and financial crime remain under development. Faster credit memorandum drafting, for example, does not establish an equivalent reduction in loan approval times. The disclosed figures nevertheless provide evidence of AI moving beyond general employee assistance into measurable changes in banking, engineering and security workflows. 4. Allica reports 99% accuracy in AI financial statement extraction UK business lender Allica Bank disclosed on 4 November that its financial statement reading agent now covers 68% of applications, with 99% accuracy in evaluation monitoring. The bank replaced an earlier two-model validation process with a single model that assigns confidence scores to extracted values. In a separate test of 100 financial statements, the revised system achieved 99.0% exact accuracy, with 0.1% of values referred for manual review. Its lending software also records human confirmations and overrides. These controls support an automated asset-finance process that reads broker submissions, requests missing information and sends qualifying cases to the bank's existing decision model. Allica had reported in April that half of the cases received through this process were decided automatically, with a mean turnaround of 12 minutes. The November update therefore provides further evidence on validation and governance rather than announcing the original deployment. Confidence thresholds, manual review and auditable decisions are important as automation extends across complex lending processes. However, the accuracy figures measure the extraction of financial information, not the correctness of lending decisions. 5. OCC sets risk-based approach for examining banks’ use of AI agents The US Office of the Comptroller of the Currency (OCC) issued a framework on 4 November to guide examiners reviewing banks’ use of AI, including software agents. Bulletin 2026-54, signed by chief national bank examiner James M. Gallagher, directs examiners to assess the authority delegated to AI systems, the level of human oversight and the potential consequences of failure. It introduces no new requirements for banks. The framework extends supervisory scrutiny to generative and agentic AI, which were not covered by the US banking agencies’ revised model risk guidance issued in April. It distinguishes between AI tools that provide internal recommendations and agents authorised to execute actions on a bank’s behalf. The latter warrant closer scrutiny where actions are difficult to reverse, particularly if human intervention is not possible before execution. This places the scope of an AI system’s authority, rather than the technology itself, at the centre of the examination approach. 6. CERT-In reports repeat attacks on lenders’ payment systems using AI-assisted tools India’s Computer Emergency Response Team (CERT-In) issued a critical advisory on 4 November after non-bank lenders and fintech firms reported further attacks on systems used to disburse funds. Its financial incident response team identified the group Crafty Falcon and reported the use of AI-assisted tools to generate and adapt attack scripts. Some firms were targeted again despite having addressed vulnerabilities flagged in earlier warnings. CERT-In’s 9 October advisory described attackers exploiting financial firms’ own payment interfaces to transfer money. The latest update details how they gained access to production systems through internet-facing test environments, using shared credentials or network connections. Attackers also made one-rupee transfers to verify that payment systems would process transactions before attempting larger amounts. The repeat incidents indicate that patching known vulnerabilities may not be sufficient to remove an attacker’s access. CERT-In has called for post-incident forensic investigations to establish whether compromised systems have been fully secured. 7. Mashreq deploys generative AI to speed up fraud investigations Mashreq and technology provider Clari5 announced on 3 November that generative AI had gone live on the Dubai bank’s fraud platform. The system consolidates information from multiple sources to help investigators examine suspicious transactions and prepare case narratives. Rivi Varghese, Clari5’s chief executive, said the software runs on the bank’s own premises, while Nitin Chugh, group chief executive of Clari5’s parent Perfios, said investigators retain responsibility for interpreting the evidence. When the partners outlined the project in May 2025, they described a tool that investigators could query in everyday language and use to draft case narratives. These functions are intended to reduce the time spent gathering information and documenting findings after an alert. The partners initially targeted a 70% reduction in investigation time. That measure, however, does not establish whether the system improves fraud detection. Assessing that would require evidence of how many confirmed cases of fraud it identifies across comparable transaction volumes. 8. Mizuho introduces central gateway to govern AI agents across banking systems Technology supplier Kong disclosed on 4 November that it had begun supplying Mizuho Financial Group with a gateway connecting AI agents and language models to banking systems under a contract signed in June. The platform provides a central point for managing access permissions and monitoring model usage and costs by application or employee. Kong said the infrastructure is designed to support Mizuho’s planned expansion to thousands of AI agents. Mizuho has already deployed AI in relationship management through RM Studio, part of its Wiz Base initiative launched in December 2025. The tool draws on previous client discussions, proposals, internal product information and public research to help relationship managers prepare presentation slides. As more business teams develop their own agents, the gateway is intended to standardise how these applications connect to banking systems and enforce access controls. This allows teams to develop task-specific tools while maintaining central oversight of their permissions and usage. 9. HKMA proposes central register and mandate checks for AI agents The Hong Kong Monetary Authority (HKMA) launched Project SAGE on 2 November and established an expert group to examine how financial institutions should verify AI agents and the authority delegated to them. A discussion paper issued alongside the initiative asks whether banks should report material AI agent deployments to a central register, recording each agent’s owner, purpose and permitted actions. The HKMA is seeking feedback by 1 February 2027. The paper distinguishes between AI agents that banks deploy in their own operations and those that act on behalf of customers. Banks can apply existing outsourcing controls when using external agents in their internal processes. However, agents acting for customers present a different challenge because banks may have limited visibility into how they are designed or operate. The HKMA’s proposals therefore focus on verifying customer mandates, limiting the instructions banks accept and ensuring they can revoke an agent’s access when necessary. 10. Gore Mutual embeds AI policy assistant in claims and underwriting workflows Canadian insurer Gore Mutual and technology provider Guidewire announced on 2 November that claims adjusters and personal lines underwriters were using ProNavigator within the insurer’s core policy and claims applications. The assistant gives staff access to underwriting rules, policy guidelines and internal procedures while they work on individual cases. Suzanne Courtlander, Gore Mutual’s chief claims officer, said the integration reduces the time staff spend locating information needed for their decisions. Previously, employees had to search separate systems and shared drives for relevant policy wording. According to Guidewire’s April product release, ProNavigator cites the insurer’s source material and applies role-based access controls to limit the information employees can retrieve. This allows claims adjusters to trace the basis for policy interpretations during reviews. The deployment changes how staff access and verify internal guidance, while responsibility for coverage decisions remains with the claims team.
