AI Security Strategy for Banks: Choosing Threats, Signals and Success Metrics

You are the product manager for a financial security product. How would you apply AI to strengthen security across banking systems? ```hint Pick the threat first Security in banking covers payment fraud, account takeover, scams that trick customers into sending money themselves, money laundering and attacks on the bank's own infrastructure. Choose one to go deep on and justify it, because each one changes who the user is, what data the model sees and how fast it has to decide. ``` ### Clarifying Questions - Is the product built inside one bank, or sold to many banks as a service that has to plug into their systems? - Which threat is in scope: card and payment fraud, account takeover, authorized-payment scams, money laundering, or cyberattacks on bank infrastructure? - Who acts on the output: the bank's fraud or security analysts, an automated decision in the payment flow, or the end customer? - Are there limits on the data we can use, such as sharing signals between banks or customer privacy rules? ### What a Strong Answer Covers - A threat chosen and justified by loss, frequency and fit with what AI does well, with the users on the bank side and the customer side named. - Concrete AI use cases matched to that threat: what the model predicts, from which signals, and at what point in the transaction flow it decides. - How the system works end to end: real-time scoring, the actions taken on a score, analyst review, and the feedback loop that turns outcomes into labels. - Metrics that balance fraud caught against false positives and customer friction, plus operational measures for analysts. - Risks specific to AI in banking: adversaries adapting, delayed labels, explainability and model-risk review, bias, and privacy. ### Follow-up Questions - A new model catches more fraud but also challenges many more legitimate payments. How do you decide whether to ship it? - Fraudsters change tactics within weeks of launch. How does the product notice that the model is going stale, and what does it do about it? - Where would you use a large language model in this product, and where would you refuse to let one make the decision?