Bank of America Puts 4,000 on AI-Found Software Flaws

Bank of America is assigning thousands of people to clean up after AI models that are now very good at finding holes in software. According to CNBC, the bank has brought in 1,000 outside contractors to repair vulnerabilities flagged by systems such as Anthropic's Mythos. Those contractors work alongside 3,000 of the bank's own employees, which brings the total to roughly 4,000 people.

CEO Brian Moynihan said the joint effort has been running since the second quarter. That makes Bank of America one of the first large US companies to say publicly how much staff it has committed to fixing flaws that AI found.

What the bank has disclosed

The figures are partial. Moynihan said Bank of America spent "a few hundred million dollars" in the third quarter. That amount covers two things: the software remediation work and the bank's anti-money-laundering requirements. He did not give a separate number for patching alone, so outsiders cannot tell how much of the spending went to vulnerability fixes.

What is clear is the scale of the staffing. Four thousand people is a large operation, even for one of the biggest banks in the US. Moynihan also offered a forecast: he expects the cost of patching to fall over time. He did not say how quickly that would happen or what would drive it down.

Where the vulnerabilities come from

The backdrop is a wave of security-focused AI releases from the frontier labs this year.

Anthropic's Mythos. In April, Anthropic made Mythos available only to a limited group of partners through an initiative called Project Glasswing. The company said the model had uncovered thousands of serious security vulnerabilities. It chose restricted access over a broad public release.

OpenAI's Daybreak. A month later, in May, OpenAI launched its own cybersecurity program, Daybreak.

At first, many critics treated these warnings and countermeasures as marketing. The labs were claiming their models were so capable at finding flaws that access had to be controlled, and some observers saw that as a convenient story.

That view has since lost ground. Two developments changed the conversation:

• Real incidents. Security breaches such as the Hugging Face hack made the risk harder to wave away.
• Independent testing. Outside evaluations confirmed that the new models really do have stronger capabilities than earlier systems.

Bank of America's disclosure adds a third data point. A heavily regulated bank does not hire 1,000 contractors for a PR exercise.

The patching problem

The story shows a gap that security teams have worried about for a while. Finding a vulnerability is only the first step. Each finding has to be confirmed, ranked by severity, fixed, tested and deployed, often in old and tightly linked systems. Banks run many of these.

When an AI model can surface flaws much faster than before, the work moves downstream. The constraint is no longer discovery. It is the number of engineers available to act on what was discovered. Bank of America's answer, at least for now, is to add people.

The same pressure has appeared elsewhere. Google recently paused an open source bug bounty after being flooded with AI-generated reports, another sign that the volume of findings can outrun the capacity to handle them.

The Bigger Picture

For readers who follow AI security, this is one of the clearest signs yet that AI vulnerability discovery has moved from lab demos to line items in corporate budgets. Earlier this year the argument was about whether the labs were overstating their models' abilities. Bank of America's disclosure suggests at least one large institution took the threat seriously enough to build a remediation workforce of about 4,000 people.

It also shows that AI is creating work here rather than removing it, at least in the short term. The models find the flaws, but humans still do most of the repair work. That fits a wider pattern in which defenders and attackers are both using the same new tools. CyrioX has reported on cases where AI hacking tools let a single attacker breach Korean banks, which helps explain why a US bank would rather fix its weak spots before someone else finds them.

There are open questions. Moynihan's expectation that costs will fall is plausible if the backlog of older flaws shrinks, or if AI starts handling more of the fixing as well as the finding. Efforts like Anthropic's free AI scanner for open source projects point toward wider access to these tools, which could help defenders but also produce more findings to process.

It is worth watching whether other large banks and enterprises publish similar numbers in their next earnings calls, and whether they separate patching costs from other spending. If more companies follow Bank of America, the true cost of AI-driven security work should become much easier to see.