HSBC plans institute to combat fraud involving AI agents

On 2 November 2026, at the opening of Hong Kong FinTech Week, HSBC announced plans for a research institute focused on artificial intelligence and emerging technologies. The bank chose the Hong Kong University of Science and Technology as its first academic partner. Among other topics, researchers and HSBC’s technical teams will study ways to thwart fraud in systems that use AI agents.

According to Reuters, HSBC chief executive Georges Elhedery aims to open the institute in the first quarter of 2027. The bank is considering locating the institute in the Northern Metropolis if it can find a suitable site there.

On 14 October 2022, HSBC and the university’s business school announced a partnership for research and training in financial technology. The planned institute will continue that training work. In addition to fraud detection, its programme covers AI safety, post-quantum cryptography and blockchain.

Reuters reports that the group’s AI training is intended for HSBC’s roughly 200,000 employees worldwide. The bank has not said how many people the Hong Kong institute will be able to accommodate.

HSBC is already running a trial separate from the fraud research planned for the institute. In remarks reported by RTHK, Georges Elhedery said the bank is testing interactions between its customers’ assistants and its own agents on simple tasks. RTHK reported no examples of completed payments or findings on customer protection from the trial.

Checking an invoice or paying it

On 2 November, the Hong Kong Monetary Authority launched its own project, SAGE, and formed an expert group. It also published a discussion paper on banks’ use of agents. Its chief executive, Eddie Yue, asked the industry whether it should create a central registry of trusted agents and establish an identification standard called “Know Your Agent”.

In its announcement launching SAGE, the Monetary Authority distinguishes an agent’s identity from the authority its user delegates to it. Consider a hypothetical example. A customer asks her assistant to check with her bank whether an invoice has already been paid. The bank could recognise the assistant when it contacts the bank. If the assistant then asked the bank to pay the invoice, that request would go beyond the customer’s instructions.

Asking the customer for approval again for each transaction would cost her some of the time she had saved. With its AP2 payment protocol, Google describes how to specify in advance which transactions are authorised and what limits apply, then link a payment order to the approved instruction. If the customer disputed a debit, a record of her request and of the order the bank received would help her show that she had only asked the assistant to check whether the invoice had been paid.

Evaluating such a control would require measuring both the proportion of fraud attempts it lets through and the proportion of legitimate requests it wrongly blocks. HSBC has not announced plans to publish those rates.