Bank of America hired 1,000 contractors to patch flaws AI found. What it means for your office

What happened

On Oct. 14, the day Bank of America reported third-quarter results, CEO Brian Moynihan told CNBC that the bank had added about 1,000 outside contractors in the second quarter to fix software weaknesses exposed by advanced AI models. He named Anthropic's Mythos as one of them. As reported by ITPro, he said those contractors are still there, on top of the 3,000 people the bank already had on that work.

The flaws weren't only in the bank's own systems. TNW reports they were in code used by both the bank and its outside vendors. On the earnings call, Moynihan said the quarter included a couple hundred million dollars or more in remediation costs, but that figure also covers fixing the bank's anti-money-laundering controls, so it isn't an AI-only number. TNW notes it hasn't independently verified the figures; they come from Moynihan's interview and the call.

The bank didn't put out a release on this. CNBC's interview page blocked our tools, so the details here come from the outlets that reported it: ITPro, TNW and Cybernews.

Why AI is creating a patch wave

Mythos is an AI model Anthropic built that is very good at finding security flaws in software. Anthropic didn't release it to the public. It gave access to a small group of companies, banks and government agencies through a program called Project Glasswing, so they could find and fix holes before attackers did.

Anthropic's own update on Glasswing, from May, shows the size of it. Partners had found more than ten thousand high- or critical-severity vulnerabilities. Cloudflare alone found 2,000 bugs, 400 of them high or critical. When Anthropic scanned more than 1,000 open-source projects, the model flagged an estimated 6,202 high- or critical-severity vulnerabilities. Outside security firms checked 1,752 of those and found 90.6% were real.

Found vs. fixed, open-source bugs (Anthropic's count)

High- or critical-severity bugs Anthropic reported to open-source maintainers, and how many were patched, as of its May 22, 2026 update. Anthropic's own figures.

1. Reported to maintainers530
2. Patched so far75

Source: Anthropic, Project Glasswing: An initial update, May 22, 2026.

Finding is now the easy part. Of the 530 high- or critical-severity bugs Anthropic had reported to maintainers at that point, 75 had been patched. Part of that was timing: Anthropic was still early in its 90-day disclosure window. But it was blunt about the cause. It said the bottleneck is people: the human time it takes to sort, report, and build and ship patches. It also noted that a recent Palo Alto Networks release had over five times as many patches as usual.

Anthropic's advice to defenders was one line: shorten your patch testing and deployment timelines. Bank of America's answer was to hire. As Moynihan put it, "Once you patch, you don't have to patch as much."

A bank can hire 1,000 people to keep up. Your company has one IT person, or an outside firm.

What this means for a mid-size company

You may think this is a big-bank problem. Banks are the targets, and they got early access to the model. But look at where the flaws were: in the bank's code and its vendors' code. The same thing is true for you. Your ERP, your accounting package, your firewall, your Microsoft 365, your shipping software, the small tools that connect them. All of it is written by someone, and AI is now reading all of it for holes, on both sides.

That means two things for your office.

More patches, on shorter notice. When vendors find and fix more bugs, they ship more updates. The Palo Alto example is one sign. Your IT person or outside IT firm will see more of them, and some will be urgent.

Less time between a fix and an attack. When a patch comes out, it tells attackers where the hole was. The same AI that helps defenders find bugs can help attackers read a patch and go after anyone who hasn't installed it. We covered that side in AI hacking tools and business apps. The window you have to install a fix is getting shorter.

For a benchmark, the federal government already sets one. CISA's directive for federal agencies requires them to fix flaws on its known-exploited list within two weeks for newer vulnerabilities. You're not a federal agency, but it's a useful yardstick. If your patches take months, that's the gap.

Where the other side has a point

Be careful with the panic, too. Moynihan didn't say how serious the flaws were, as Cybernews notes. And he raised a fair point himself: the companies whose AI finds the flaws are also selling the services to fix them. He said the bank has to keep in mind that the people building the software are also the ones selling it the services, per ITPro. Anthropic's numbers are Anthropic's own, even with outside firms checking a sample.

It's also true that most break-ins at companies your size don't need a fancy new bug. They use stolen passwords, missing multi-factor login, and old flaws that were never patched. The AI patch wave doesn't change those basics. It makes skipping them more expensive.

What to do this month

1. List what you run. Every piece of software and every device that touches your network, who makes it, and who updates it. You can't patch what you don't know you have.
2. Set a patch window and keep it. A fixed time each month for routine updates, and a faster rule for urgent ones, such as within two weeks for anything on CISA's known exploited vulnerabilities list.
3. Start with what faces the internet. Firewalls, VPNs, remote access tools and email come first. Those are what attackers reach first.
4. Ask your vendors three questions. How do you find security flaws in your own code? How fast do you ship fixes for critical ones? How will you tell us?
5. Ask your IT provider for a number. How many systems are behind on updates today, and how old is the oldest missing patch?
6. Don't forget the small tools. The scripts, add-ons and connectors that tie your systems together are software too. Someone has to own them.

None of this needs 1,000 contractors. It needs someone to own it, a list, and a calendar. For more on the AI side of security, see our guides on attackers going after your AI tools and the cheap devices on your network, and our security page.