In an AI world, compromised security will lead to fraud: Matthew Driver

Mastercard says stronger cybersecurity, agent verification and consent frameworks will be critical as AI agents begin making purchases on behalf of consumers

Matthew Driver, executive vice-president for services in the Asia Pacific at Mastercard, believes companies must strengthen cybersecurity amid a complex software environment and the rise of agentic commerce — where autonomous artificial intelligence (AI) agents independently research and execute online purchases for customers. Driver, in an interview with Ajinkya Kawale and Shivani Shinde in Mumbai, explains the global card network’s priorities in an AI world. Edited excerpts:  Would you still control the customer’s experience in agentic commerce? 

Ultimately, the idea of agentic commerce is somebody is doing something on your behalf. The critical thing to make that happen is you need to go from KYC (Know Your Customer) and KYB (business) to KYA (agent). I need to think about the consent or instructions that the customer is giving the agent, verifying the agent, making sure where it has come from, what kind of data it was trained on. I’ve got to do consent tokens. I need verifiable intent and the ability to tokenise all those instructions and essentially establish a chain of provenance through the ecosystem. I need to be assured if somebody makes a mistake, or my agent or the merchant makes a mistake, I have a recourse. 

What will be the major use cases in an agentic ecosystem? 

People inherently drive for convenience. Everyone is looking for an AI summary. If you are going to be participating in this, whether you are a bank or a merchant, you have to be discoverable to a machine. So, you need to think about whether I need an agent to be able to populate, utilise an application programming interface to get inventory, pricing, stock keeping units, and generative engine optimisation? One of the services we do is help with discoverability. We look at the issuing side and see if one needs tokenisation, passkeys, biometrics, so then you’re going to be able to be supported during checkout. Price watch (tracking item prices) is a big use case. Curate and discover is another. For instance, I want to go to a concert, rather than chasing tickets, an agent can do it. In talking about agents, you have got to be able to experiment in spaces where there is tolerance for making mistakes. That’s where you will see the first use cases starting. 

Claude Mythos, Anthropic’s advanced AI model, has shown how vulnerabilities can be accessed and potentially exploited. How are you keeping pace with it? 

We are aware that advances in AI and compute have advantages for companies but they also naturally establish threat vectors because effectively they’re democratising access to very powerful tools. We continue to be very vigilant in terms of what we do in our own network. Mastercard has an incredibly secure payment network. We also have a relationship with Anthropic and so I won’t talk about the details of that. We are trusted because Mastercard has not suffered any significant compromise in any way. More broadly, we’re seeing a degree of inflation of cybersecurity and scams and fraud, in the sense that an upstream compromise might lead to a downstream fraud. We’re trying to make sure that we’re encouraging public-private partnerships to ensure we're collaborating and sharing information. We want to utilise capabilities like threat intelligence to give clients insights, maybe around patching prioritisation. 

With AI models like Claude Mythos, do you see cybersecurity spending go up? 

It’s a bit of a speculative question. I don’t really have a good answer to it because token cost is going down but the usage is massively going up. We’ve got to harness AI in the right kind of way. We’re probably going to see that downstream there’s going to be much more usage of small language models. The reality also is that there’s a lot of focus on return on investment now and effective use of tokens. 

How do you ensure AI agents are robust in a transaction environment? 

We’ve got 18 to 19 different variables that we monitor. I need to make sure that I have got a registry of agents at manufacturers. Then, I am monitoring the behaviour of those agents over time. Is there a model drift? What happens to toxicity and behaviour? Do we see more disputes as a result of a particular agent? Then, I have got to be able to have network governance and naturally it is in our best interest to set some standards. If you violate performance requirements, we will subject you to review, potentially penalties or otherwise.