Webinar Highlights: Financial institutions and AI-driven cyber attacks - Part 1 of 3

Hello everybody, I'm Bruno Ganchard. I'm your moderator today. A few words about me. I spent the last 18 years at UBS, a Swiss bank, in the IT area, quality assurance, testing, compliance, and more recently AI governance. I would like to start with some context. In April, Anthropic announced the Mythos, and within a couple of weeks was able to identify more than 10,000 vulnerabilities. Anthropic decided instead of commercial release to create Project Glasswing, where a limited number of companies were invited. Only US companies, including the hyperscalers Google, Amazon, Microsoft, they were able to test these models and show their own defense. And in parallel to that, OpenAI released ChatGPT-5.5 Cyber and gave also the opportunity to a limited number of global companies, including Europe, to test their model. Last Tuesday, Anthropic announced that they extended their project Glass Ring to more companies, especially in Europe. And there is some discussion with the European Union about the access conditions. With all these events happening, what is the main impact for the financial institutions? And for us, what should we do? And we are going to discuss that with 4 lenses. The first one would be trust, all the trust we are building with our clients. The second one is the continuity, how we can ensure business continuity when we have new threats. Third one will be accountability, or what should be the role played by the regulators in that game. And the last one is what next, what other action we need to take now to make sure that we are prepared with what is coming. To tackle them, I have a couple of speakers around me. We have Andres Kucharavy. Andrei is a co-founder of Swelio. He's also a professor in machine learning and cybersecurity. We have Daniela, founder of DNYC. Daniela works closely with fintech companies in the ecosystem, so she'll bring all experience in that area. And I'm Colin Perard, co-founder of Stratage and work as a consultant. And we'll start with you, Andrej. What's happening now? And we have OpenAI, we have also Mythos coming into the game. Are these models a game changer or it is a trend that was already there in the past? How do you see that from a technical perspective?

Yeah, to give some more background, I have been researching the use of generative AI in cybersecurity. With the Swiss Cyber Defense Campus since 2020, so close to 6 years now. One of the things that we were already seeing back then is that even today, outdated models such as GPT-3 were really good at finding vulnerabilities in hands of competent people. So for instance, there was a researcher at Pennsylvania that in 2021 used GPT-3 to find over 200 vulnerabilities in open source projects, whereas the best specialized tools back at the time could only identify about 100. So we knew that LLMs would have capabilities for finding vulnerabilities. However, there's 2 important stages that were present until recently. First one is that the operator of those tools needed to be somebody who was quite competent and knew where to point them, how to operate them, and how to transform this information about the vulnerability into an attack. And second was weaponization. Once you have a vulnerability, you need to figure out how you're going to use it. Maybe it's well hidden, you need to access to it directly. Maybe it doesn't really give you that much capabilities and you need to combine it with somebody else. That's where LLMs that were able to code came into place and really started improving things starting from 2022, but especially '23. And the transition we started observing at the end of '24 and '25 is that models that were fairly advanced, but not only from Anthropic OpenAI, also code assistance model from China, from Mistral, were able to start putting together all those ideas with a little bit of guidance from people who still needed to be competent, but that could really accelerate things. What's happening starting from the end of 2025 is that Claude Opus and to some degree GPT-5 were able to automate this putting together through their reasoning chains and to a substantial degree start automating things. And already in autumn 2025, what we started seeing is that in CTF competition, capture the flags, which is kind of games for hackers where they try to attack on a machine that they know is vulnerable, teams are starting to rely heavily on cloud ops providing them capabilities of solving challenges that were prepared for one day in a couple of hours. The real change that Mythos brought is that it was optimized for an attack, so they had an offensive function added. into it. And there's a lot of debate in the AI safety community whether it was a good idea, because while we now have very strong automated attack capabilities— that's not only due to Mythos, GPT-5.5 has the same capabilities, Microsoft has their own models that's able to do the same thing— in absence of the tools to defend against them, is really that good of an idea? Because this can be cloned. Mythos was already stolen from Anthropic. So bad guys have the tools to attack fast, to attack strong. even if they are not very competent. And as defenders, we have nowhere near the toolbox to respond to it. And we currently have this gap between attackers and defenders that Mithras and the hype around it brought to public attention, but that was already there from late 2025. And that is likely to get worse in the upcoming 18 to 20 months, in my opinion, before it gets better.

Thank you for this very good explanation. So that means that a new threat's coming in and the financial institutions were always like a prime target for that.

I do think, Daniela, that As you said, financial services have always been targeted by thieves because they store money or because they facilitate the transfer of money. But what we see nowadays is that financial services actually are the connective tissue of the economic environment. Malicious actors, sometimes state-sponsored, might use attacks to financial institutions, not just to steal money, but to disrupt entire economies. And they have the potential to do that. We tend to differentiate between tier 1 banks and financial institutions and tier 2. We've seen with Anthropic, they try to favor the tier 1s, but actually, if we look at small-medium enterprises, enterprises, which are the thousands of companies that hold economies together. And they're in most countries, the propelling force of economies. They actually rely on the tier 2 financial institutions as well, the cooperative banks. And so they are smaller institutions. There are more of them. They are perceived to have less defense mechanisms than the tier 1s do. So this is really the topic that everyone is concerned about. The central banks are concerned about that. The regulators are concerned about that. And this is because they want to maintain trust in the system. any of those malicious actors were to succeed in one of their attacks, it's not so much the damage that they would cause to an individual institution, but the damage that they could cause to the trust in the system. And this is what we need to prevent.

Yeah, as you said, trust is key here and it's not an IT topic, it's really a global topic. So I assume, Caroline, you are also interacting with a lot of boards. What are they discussing currently and what are the fears they express at this stage?

Thank you, Bruno. I thought we might discuss about 3. One is definitely how we ensure that the AI models, the new technologies are going to be compliant with the regulation because the regulation is evolving. We rely on different regulation in every jurisdiction. Some are more sector-based, some are very stringent rules, some are relying on existing legislation. And then we know that in Europe, Switzerland, and other countries, regulation is on the making with a lot of guidelines to understand how the technology works and what could be the risk. So definitely, is it compliant? And the second question, how do we audit and we document AI decision-making for regulators? And so very important for financial institutions, they need to document and ensure that they do the proper audit. And if we take the example of our AI-driven solution for credit limits, if this solution is including biases, then the result might be discriminatory. And then the company will have to show that some measures have been taken to avoid this bias. So it's really about how do we audit and document before the harm is done? And the last question they ask is, how do we trust this solution with the critical financial decision? And how do we mitigate the systemic risk? Mitigation is definitely a topic that is coming up involving a lot of technological knowledge And then at the end of the day, is it worth solving the problem with this tool while we need to put a lot of measures to maybe mitigate the risk, to do some pre-testing, verification, et cetera? So the tool seems promising, but they impose a lot of extra work to make sure that the financial institutions are doing it well. So I would say the main fear they have is fines, loss of reputation and trust. And then in the financial sector, loss of licensing, the loss of the ability to operate. And that's a very high burden that they need to take upon.

Okay. Thank you. Compliance risk is really a big one. A key topic for all institutions. I'm coming to you, Daniela, about your experience. And we know that fintechs are more agile, can move quicker than the tier 1 banks. But on the other side, I think they may have less robust infrastructure or they have less capacity to protect themselves. See this as a threat, as an opportunity, or both?

It's a bit of both. And I'll focus particularly on the European fintechs because that's what I'm closest to. I think when it comes to cybersecurity, there is an opportunity because particularly the largest fintechs in Europe are actually perceived to be more agile and potentially more resilient. to cyberattacks. I mean, this is all in principle, and I stress that it's a perception, but it could be an opportunity to win market share versus competitors. At the same time, it is a threat because fintechs do not operate in a silo. They do not operate in a parallel world. They are interconnected. Oftentimes they are resellers of services or products by very traditional players. And because they are part of this, an attack to the ecosystem is also a threat to the fintechs. The fintechs are fully aware of it, and that's why this enters in their risk assessments.