Bank of America has put numbers on what it costs to defend against AI that finds software flaws. CEO Brian Moynihan told CNBC on October 14, 2026 that the bank added 1,000 contractors in the second quarter to help patch vulnerabilities exposed by advanced models including Anthropic's Mythos, on top of about 3,000 people already working in that area. In CNBC's report, the bank says those contractors are still on the job.
It is one of the first times a major US company has specified the manpower and dollars spent protecting systems against Mythos, which CNBC says was released in April to a small group of banks, technology firms and government agencies. This post covers what is confirmed, what the cost figure does and does not mean, and the practical lessons for any team whose code an AI model could now scan faster than a human can fix it.
Quick reference: what was said and what it means
| Question | Answer |
|---|---|
| Who said it | Brian Moynihan, CEO of Bank of America, to CNBC after the third-quarter earnings release (October 14, 2026) |
| Staffing | 1,000 contractors added in Q2, still on site, alongside about 3,000 internal staff in that area |
| What they patch | Proprietary code and vendor software with flaws exposed by frontier models including Mythos |
| Third-quarter spend | A "couple hundred million plus" on software remediation and anti-money-laundering compliance combined |
| Outlook | Moynihan expects remediation costs to decline as patching catches up |
| Related AI savings | About 130 internal AI projects; savings rising from $300 million this year to $700 million in 2027 |
| Caveat | The dollar figure mixes two cost lines; the Mythos share is not broken out |
What Moynihan actually said
Moynihan described the response as a rush. Earlier this year, he said, the company moved to strengthen a wide spectrum of internal and vendor code using thousands of employees and outside personnel. His account of the thinking is direct: "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?'" He then gave the headline number: "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 also framed the dilemma every large buyer of software faces: "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." In other words, the vendors whose products carry vulnerabilities are also selling the tools and services to find and fix them.
Asked about cost, Moynihan told analysts earlier in the day that the bank spent a "couple hundred million plus" in the third quarter on software remediation and on regulatory issues related to anti-money-laundering compliance. CNBC notes bank shares fell as much as 8% on the day, and that Moynihan sought to reassure investors the increase was temporary: "Once you patch, you don't have to patch as much... The pipeline moves through."
Why the cost figure is easy to misread
Several outlets have repeated "BofA spent hundreds of millions on Mythos." That is not what the CEO said. The number covers two separate things: software remediation and anti-money-laundering compliance work. One of those, the AML spend, has its own story; CNBC's video on the same page is titled around a $1.2 billion fine and the bank's plan to remediate. So the honest summary is that the bank's combined third-quarter spend on those two areas ran to a couple of hundred million dollars, and no source we found separates how much of it is Mythos patching.
The staffing figure, by contrast, is clean and quoted directly: 1,000 contractors on top of 3,000 staff. At typical contractor rates that implies a large cost, but we will not guess a number, because the contract terms are not public. If you are citing this story, quote the headcount and say the dollar figure is combined.
How does this fit with what we knew about Mythos and banks?
Banks were among the first to feel Mythos. Anthropic's model, announced in April for a limited set of defenders, was pitched as able to find vulnerabilities at a scale human teams cannot match; our background piece on Claude Mythos Preview and Project Glasswing covers the launch. Reuters reported in May, per search summaries we reviewed, that a handful of the largest US lenders had access, that the tool was surfacing from several hundred to thousands of mostly low-to-moderate vulnerabilities, and that banks were cutting patch times from weeks to days; PYMNTS summarized that reporting. Moynihan's remarks are the first time we have seen the workforce side of that pattern spelled out.
Since then Anthropic has built out a structured program for who gets to use these capabilities. Our write-up of the three-tier cyber verification program explains how defenders, red teams and specialized users are vetted, and the Mythos 5 security scans and defender-advantage fund covers the money Anthropic put behind defenders. Rival tools are catching up too: see how Zhipu's model matches Mythos on security bugs and how OpenAI's GPT-5.5 compares for cyber defenders.
The defender's dilemma, in numbers
Three dynamics explain why a bank would hire 1,000 people for patching:
- Discovery is now cheaper than remediation. An AI model can surface hundreds or thousands of candidate flaws in the time it used to take a team to triage a handful. The bottleneck moves to the humans who must reproduce, prioritize, fix, test and deploy each patch, and in a bank every change goes through change control.
- Vendor code is half the problem. Moynihan specifically mentions vendor software. A bank cannot patch what it does not control, so it must pressure suppliers, wrap legacy systems, or replace them.
- Attackers get the same capability. Defensive patching is a race against the day an adversary runs a comparable model against the same code. That is why the bank treats the spend as risk reduction rather than optional hygiene.
The dual-edged nature is the point CNBC makes: AI can save money through productivity, with Bank of America projecting savings from its roughly 130 internal AI projects rising from $300 million this year to $700 million in 2027, while the same class of models forces a costly defensive scramble. CNBC adds that the bank's headcount fell by about 3,400 over the past year as it cut roughly 5% of managers, but it is unclear how much of that is attributable to AI.
What this means for teams outside banking
You do not need a bank's budget to apply the lessons.
- Plan for volume, not just severity. If an AI scan returns thousands of low and medium findings, you need a triage policy: auto-close the unreachable, batch the similar, escalate the exploitable.
- Inventory your vendor code. Know which third-party components sit in your critical paths and what your contractual right to demand patches is.
- Speed up the pipeline. Faster tests and deployment are now a security control. A patch that takes six weeks to ship is a six-week window.
- Use the free tools you qualify for. Anthropic's free OSS scanner for open source maintainers is one example for projects that do not have a security team.
- Put guardrails on your own agents. Patching increasingly involves AI agents with repository and production access. AgentBeam, the agent security platform from the explainx.ai team, stops AI agents before they take dangerous actions, which matters when the agent doing the patching has write access to the code you are trying to protect.
What to watch next
Other large banks report third-quarter results over the coming days and weeks, and analysts will likely ask the same question Bank of America's executives faced: how much of the cost increase is AI-driven remediation, and when does it fall? If peers disclose comparable headcount or spend, the first benchmark for the cost of defending against AI-found vulnerabilities will exist. Regulators are another thing to follow, since supervisors have an interest in whether large financial firms can patch quickly enough, and in whether they depend on a handful of AI providers for the discovery step.
For the AI labs, the story is also a demand signal. A vetted-access model that creates work for thousands of contractors at one customer is evidence that AI security capability has moved from demo to line item. The open question is the balance: if automated fixing improves as quickly as automated finding, the contractor count could shrink, and that is the trend a reader should look for in next quarter's numbers.
What we could not verify
CNBC's report rests on Moynihan's own statements; the bank has not published a breakdown of contractor cost, the number of vulnerabilities found, or how many were critical. We also could not confirm which AI tools or access tier the bank uses internally, so "Mythos" here means the class of frontier models the CEO described rather than a specific deployment. A Bank of America regulatory filing or the earnings call transcript is the place to look for any more granular numbers.
Related reading
- Claude Mythos Preview and Project Glasswing
- Anthropic's three-tier cyber verification program
- Mythos 5 security scans and the defender-advantage fund
- Anthropic's free OSS scanner for open source maintainers
- Zhipu AI matches Mythos on security bugs
- OpenAI GPT-5.5 for cyber defenders vs Mythos
- VulnCheck: AI-found bugs and exploit rates
Details are accurate as of October 14, 2026 and based on CNBC's interview report with Bank of America's CEO.