Bank of America put 4,000 people on patching what AI found
Two stories about the bill that arrives once a model tells you the truth — one bank paying it, one company selling a device that listens in.
Bank of America has 1,000 outside contractors working alongside 3,000 internal staff to patch software flaws that frontier AI models exposed in its own code and its vendors' code. CEO Brian Moynihan told CNBC the contractors came on in the second quarter and are still on the job, and that software remediation combined with anti-money-laundering compliance added "a couple hundred million plus" to third-quarter costs — the bank did not separate patching from compliance. He expects the spend to taper: "Once you patch, you don't have to patch as much. … The pipeline moves through."
That is about as concrete as the defensive side of AI security has gotten. It is also, as far as we can tell, the first time a major US company has put a headcount and a dollar figure on cleaning up what an offensive model found. The model in question is Anthropic's Mythos, released in April to a small set of banks, technology firms and government agencies, which the lab said had surfaced thousands of serious vulnerabilities. We looked at the other half of that story this week — Chainguard puts Claude Mythos on vulnerability hunting.
The interesting part is the shape of the ledger. In the same earnings call, Moynihan said Bank of America's roughly 130 internal AI projects are on track to deliver $300 million of savings this year and $700 million in 2027, and that headcount fell by about 3,400 over the past year as the bank cut roughly 5 percent of its managers — with no claim that AI caused the cuts. Savings arrive as a projection; the patching bill arrives as an invoice. That asymmetry is the honest state of enterprise AI right now, and it is worth watching whether the next bank to disclose numbers reports the same ratio.
Amazon's revamped Bee wearables launch with an $11.99 monthly subscription and an always-on microphone. The Frontier is a wristband at $75 with up to seven days of battery; the Eclipse is a credit-card-sized magnetic snap-on for your phone at $100 with 21 days or more. Both feed transcripts into templates — lectures, care plans, brainstorming — pull in Gmail, Google Calendar, Outlook, Apple HealthKit and Alexa+, and push shopping suggestions from Amazon.com. Amazon's safeguards are specific: no audio is ever stored, processed speech is deleted after transcription, and transcripts are encrypted with keys the device creates and the customer controls. What none of that settles is the other person in the conversation, who never agreed to be indexed — a question we ran into this year with OPPO's 499-yuan AI pendant records your day from your collar.
What to watch: whether Bank of America's remediation line actually shrinks next quarter — that is the closest thing we have to a price for the post-Mythos world.
Would you pay a monthly subscription for a device that remembers your conversations? Tell us in the comments.
