- Key insight: The threat that cybercriminals could use Anthropic's Claude Mythos model to exploit software vulnerabilities at banks or their vendors is real, experts say.
- Expert quote: "People are freaking out. I think that the Defense Department is clamping down and going to stop any release of similar tools." –Alexandra Mousavizadeh, CEO of Evident
- Forward look: Banks are likely to hire more cybersecurity experts and clamp down on third-party software security.
The initial excitement and anxiety over Anthropic's Claude Mythos Preview model has settled a bit. But banks are still left with a pressing need to deal with the threat.
Mythos is a large language model (LLM), AI-based program that is designed to find flaws in software code. The concept is not new; vulnerability scanning software has been around for years. But Mythos accelerated the time frame exponentially. Flaws that used to take years to surface now can be exposed in hours. LLMs from OpenAI, Google and others are expected to have the same capability in the near future, creating even more opportunities for bad code to be exposed. The risk of course is that hackers could take advantage of this vulnerability.
The
"And people are freaking out. I think that the Defense Department is clamping down and going to stop any release of similar tools."
In early April, Anthropic announced that its new model can autonomously find and weaponize software vulnerabilities, turning them into working exploits without expert guidance. It had found small flaws in software that thousands of software developers working on those systems had failed to see.
The banking industry risks misjudging where it hits hardest, according to Sumeet Chabria, CEO of consultancy ThoughtLinks. Mythos can find zero-day vulnerabilities, in other words software flaws that haven't been identified before, and there's the possibility that these vulnerabilities will be exploited before companies have a chance to fix them, a remediation gap.
"Mythos has not just collapsed the discovery timeline from years to hours, it has inverted the security bottleneck entirely," he said. "If fewer than 1% of the zero-days surfaced by Mythos get remediated quickly using current methods, we have not solved a security problem. We have created an unprecedented remediation gap. Unless firms fundamentally change how they redesign their security frameworks and respond, that gap will widen."
Publicly downplaying risks, privately concerned
In public, bank executives and others
"I was speaking to some of the big banks about Mythos, and it's a real problem, because this is a technology that can completely uncover loopholes in security that have been difficult to uncover with tools that were just a month old," she told American Banker in an interview. "When you have the big banks saying it's going to take us a year to create the updates that are needed in order to prevent damage if anyone outside the bank got access to this tool, I believe them."
In public comments, banks have projected confidence about Mythos.
Ted Pick, CEO of Morgan Stanley, said during his first-quarter earnings call that his bank is working with the beta version of Claude Mythos.
"We are looking at different places inside of infrastructure where there'll just be continuous improvement," Pick said. "And that's going to go on with the firms that have the history that we have of cybersecurity infrastructure as the number one priority. This is not a new phenomenon."
During JPMorganChase's call, CEO Jamie Dimon said, "Of course, we read about Mythos, which we're testing now and looking at. And it does create additional vulnerabilities, and maybe down the road, better ways to strengthen yourself too."
Protecting the bank from cyber risks is "a full-time job and we're doing it all the time," Dimon said. "And remember, you have cyber criminals, you have cyber states, you have cyber everywhere, and that's why you have to be quite careful. So, I'd say the banks in total are rather well-protected. That doesn't mean everything that banks rely on is that well-protected."
Joining Project Glasswing
In early April, Anthropic
Some experts say more banks should join Project Glasswing.
"I'm a believer of, the more voices you have, different experiences, the more creative the group becomes," Alenka Grealish, emerging tech intelligence and advisory services lead at Celent, told American Banker. "And to bring the Europeans in is really important. Anthropic is heading in that direction. They're bringing in U.K. banks. They recognize that they need to get more experts at the table and find it's a utility level where competitors do have to work well together."
Banks that can't join Project Glasswing can still get information about its findings, according to ThoughtLinks' Chabria. They can ensure their cloud, security and infrastructure vendors "are legally and technically structured to deliver upstream, proactive alerts rather than just patches," he said.
Smaller banks have the advantage of nimbleness.
"Bankwell doesn't compete with JPMorgan on security spend," Ryan Hildebrand, chief innovation officer at Bankwell Bank, told American Banker. "We compete on speed of decision-making. A model that finds vulnerabilities in three hours instead of three months rewards a bank that can also act in three hours.
"Mythos isn't the story," he said. "I'd say the story is which institutions can move fast enough to use it."
Hiring cybersecurity talent
Bankwell is evaluating AI-assisted security analysis where it fits and bringing in security leadership that orients toward speed.
"Hiring matters more than any single tool," Hildebrand said.
The fact that bank CEOs are aware of Mythos means there will probably be "some very vigorous galvanizing and shifting of budgets towards chief information security officers," Grealish said.
Banks will be hiring cybersecurity experts, "because banks over the past decade or plus have accumulated some technical debt in that while they are migrating to a modern infrastructure running more apps in the cloud," Grealish said. "There's still legacy, and that legacy is messy and likely vulnerable, so their technical debt is going to cause them to have heightened concern compared to younger AI-first industries that were built on a lighter tech stack that is a lot more transparent."
Mainframe and older core-system break-ins are rare. But, as Grealish pointed out, advanced AI models like Mythos can read COBOL, an older coding language upon which many of those systems were built, and quickly figure out how to compromise it.
Vetting third-party software even more closely than usual
Dimon spoke during his earnings call of the need to test new software before deploying it, especially code coming from third-party vendors.
"Did you ask them to do certain things to protect their company?" he said. "How do you protect your data? How do you protect your networks, your routers, your hardware, changing your passcodes? Doing all those things right dramatically reduces the risk."
He suggested banks typically do a good job of this. "You've seen a lot of banks, they haven't had some of those risks like ransomware and things like that, at least not that I know of," Dimon said.
Chief Financial Officer Jeremy Barnum added, "Yeah. Knock on wood."
Grealish noted that software as a service has helped banks keep costs down.
"But what do they not know about their software-as-a-service providers on the security front?" she said. "I think it's causing some legitimate alarms, because not only is there this sense of imminent threat, but also there's acknowledgement that before they were looking for the flaws, they were investing in what they call penetration tests. Now, discovery [using tools like Mythos] is going to be easy."
Organizations need to put guardrails around third-party software, she said. Sometimes employees install third-party plug-ins and software and the company is not even aware of it.
Tal Kollender, co-founder, CEO and CISO of Remedio, an endpoint security provider, also believes a lot of the risk lies in the possibility that hackers could use Mythos to break into third-party software programs.
"Everyone is connected to the cloud, so getting from the cloud to the crown jewels is not as hard anymore," she said.
According to Hildebrand, the bank attack surface lives at the core processors, digital banking providers and origination platforms.
"Those vendors are where the industry should be focused, and conversations are barely started," he said.
The real danger sits in third- and fourth-party vendor systems or proprietary systems running on aging infrastructure, Chabria said.
"Threat actors will use AI to hunt for cascading vulnerabilities across these partners," he said. "If a vendor cannot demonstrate they are using frontier AI to audit their own code at scale, they are an active liability. Procurement decisions are now security decisions."
Fast patching
The biggest challenge of dealing with the risk of Mythos is, of course, getting ahead of it and patching any flaws in a company's software that it might be able to exploit.
"I would expect those that are in the forefront will start figuring out how they can scale patching, and to what degree they need humans plus LLMs to scale that, because I'd imagine they're discovering hundreds of zero-day vulnerabilities just in this testing period," Grealish said. "How do you triage? How do you patch? There's going to have to be some budget shuffling."
Chabria recommends using AI to fix the vulnerabilities AI finds.
"Compress your patch-to-deploy cycle from days to minutes wherever automation allows," he said. "Institutions that continue to triage at human speed will fall further behind with every cycle. Budget-wise, boards should not plan for a 10% incremental bump. They should prepare for a potential 2x step-up dedicated specifically to machine-speed remediation. The instinct to solve this by hiring more security analysts is exactly wrong. You need builders. You need engineers who can architect and deploy automated remediation at scale. Human-speed defense does not survive machine-speed."
Sometimes the very patches issued to fix software vulnerabilities have vulnerabilities, Kollender pointed out.
"People say, Oh, you need to patch fast," Kollender said. "And if you patch fast, sometimes you don't know the impact, and then some other things happen. Then you need to undo, then you need to wait for Microsoft. Then you are still exposed. It's always the circle of, trust but verify. And sometimes there is no time to verify."
Banks should activate sector partnerships with urgency, Chabria said. They should automate threat intelligence feeds from FS-ISAC into their infrastructure and align with the FSSCC for systemic resilience.
"The truth of the matter is, there's hundreds of pretty significant financial institutions that are not well protected, that don't have an ability to inoculate themselves against this," Bijan Sanii, president and CEO of engineering firm Inetco, told American Banker.
