Bank of America added 1,000 contractors to patch software flaws exposed by Anthropic’s Mythos

What did Bank of America disclose?

Bank of America has put the first concrete headcount on what it costs a large company to keep up with AI models that find software bugs. In an interview with CNBC on 14 October 2026, after the bank released its third-quarter results, Brian Moynihan said Bank of America rushed earlier this year to shore up a wide range of its internal and vendor code.

“We put on 1,000 contractors to help us do the patching work in the second quarter, and those people remain here today, helping us, on top of the 3,000 people we have in that area,” he told CNBC. Asked why, he pointed to the arrival of the latest models: “That risk is real risk, and that’s why when the frontier-type models came out. … A lot of us sat down and said, ‘How do we figure out how to get through this quickly?’”

CNBC described the comments as some of the first by a major US company to specify the people and money spent protecting software from Mythos. Moynihan named Anthropic’s model as one of the advanced systems behind the work. CNBC’s report does not say whether the contractors use Mythos themselves, and the bank has not published a list of what they fixed.

How much the patching is costing

The cost came up on the earnings call when Wells Fargo analyst Mike Mayo asked how far AI could take the bank. Moynihan broke down the year-on-year rise in expenses. The total base was up $1.2 billion while headcount was down about 3,000. Around $450 million of the increase was wealth-management incentives tied to revenue growth and about $300 million was in the markets business. Of the roughly $400 million left, he said, “a couple hundred million plus in the third quarter were two things, remediation for the AML and remediation for the Mythos implementation.”

That figure covers both programmes, so the call does not give a separate number for AI-driven patching. Moynihan added that “we expect that the Mythos remediation costs run down because we added 1,000 extra staff to work on that, not in the headcount but in consultants.” In other words, the 1,000 people sit outside the bank’s reported employee count, and their cost shows up as an expense line rather than as hiring.

The quarter was also weighed down by a $1.2 billion regulatory settlement charge. The bank’s results release reported net income of $6.9 billion, or 91 cents a share, and adjusted net income of $8.1 billion excluding the charge. Quartz reported that the anti-money-laundering costs follow a $1.2 billion penalty from the Treasury Department’s Financial Crimes Enforcement Network, and that the shares fell as much as 8% on the day.

Why a bank needed 4,000 people to patch software

Anthropic introduced Claude Mythos Preview in April through Project Glasswing, which gave launch partners including Amazon Web Services, Apple, Microsoft, Google and JPMorganChase, early access for defensive security work. Anthropic says it extended access to more than 40 further organisations that build or maintain critical software and committed up to $100 million in usage credits.

JPMorganChase was the only bank named among the launch partners. Bloomberg reported on 10 April, citing people familiar with the matter, that Bank of America, Goldman Sachs, Citigroup and Morgan Stanley were also testing the model internally. By the summer, Anthropic said, Glasswing partners had found at least 129,000 verified vulnerabilities.

Finding bugs faster creates a second problem: someone has to fix them. In April, American Banker quoted Sumeet Chabria of the consultancy ThoughtLinks warning of “an unprecedented remediation gap” if firms kept triaging at human speed, and Celent’s Alenka Grealish predicting that banks testing the model were discovering hundreds of zero-day flaws and would need to work out how to triage and patch them at scale. Bank of America’s answer, at least for now, was to add people.

The vendor problem Moynihan described

Much of a bank’s exposure is in software it buys rather than writes. Moynihan said the patching covered both proprietary code and vendor software, and he described an awkward dependency. “We’ve got to stay ahead of them, and we’ve got to be mindful that the people doing the software development are the people trying to sell us the services,” he told CNBC.

That echoes what bank security people said earlier in the year. In the American Banker piece, Bankwell Bank’s Ryan Hildebrand said the real attack surface sits with core processors, digital banking providers and loan origination platforms, and Chabria argued that a vendor unable to show it uses frontier AI to audit its own code is “an active liability.” JPMorgan’s Jamie Dimon said on his bank’s first-quarter call that Mythos “does create additional vulnerabilities, and maybe down the road, better ways to strengthen yourself too.”

Quartz, citing earlier reporting, said US banks were finishing fixes within days by May that they might once have deferred for weeks, with the number of flaws found ranging from several hundred to thousands, mostly rated low to moderate in severity. The same pressure is changing how banks treat AI agents more broadly: this week JPMorgan said it will block outside AI agents from acting inside Chase accounts.

AI savings on the other side of the ledger

Moynihan framed the patching bill as temporary. “Once you patch, you don’t have to patch as much. … The pipeline moves through,” he told CNBC. On the call he set it against the savings the bank expects from its own AI work: about 130 implemented projects, each with a payback, which he said return two to three times their cost over time.

He put the savings from those projects at $250 million to $300 million this year, rising to $700 million next year. One of the biggest areas is software development itself, where the bank spends about $4 billion a year on change; Moynihan said it is now 10% more efficient, “so next year we get 10% more work for the same cost.” The results release says more than 200,000 employees use AI tools and the bank has more than 300 approved use cases.

Headcount fell by nearly 3,500 from a year earlier, chief financial officer Alastair Borthwick said, and CNBC reported that the bank cut roughly 5% of its managers. It is not clear how much of that reduction is due to AI. Other large lenders are making similar bets; Barclays plans to give Claude Code to half its developers by the end of the year.

What the numbers do not show

The disclosure leaves several gaps. Bank of America has not said how many vulnerabilities Mythos or other models surfaced in its systems, how severe they were, or whether any had been exploited. The “couple hundred million plus” is a combined figure for two unrelated remediation programmes, so neither programme’s cost can be read from it. And the 3,000 employees were described as working “in that area,” not as a team created for Mythos.

It is also not a claim that Mythos found every flaw being fixed. Moynihan spoke of software vulnerabilities exposed by advanced models “including” Mythos, and the patching extends to vendor code that other banks and companies also run. The figures are management’s own, given in an interview and on an earnings call, rather than audited line items.

What to watch next

The test of Moynihan’s forecast is whether the remediation line shrinks in the fourth quarter and whether the 1,000 contractors leave, and whether other lenders disclose comparable numbers. On the supply side, Anthropic has been turning Glasswing into broader programmes, including a critical-infrastructure defence effort that it says is aimed at the slow step between finding a bug and fixing it.

For other companies, the lesson from Bank of America is that the cost of AI-found bugs lands on the people who own the software, and on their suppliers. A bank with 3,000 people already in the area still needed a third more to keep up.