The neural network has found so many weaknesses from Bank of America that the bank had to dramatically increase the human team for correction. Bank of America chief Brian Moynihan told CNBC that in the second quarter, the bank had attracted 1,000 external contractors to patch vulnerabilities identified by advanced AI models, including Anthropic’s Claude Mythos Preview.
The contractors joined about 3000 employees who were already doing similar work. In mid-October, an additional thousand specialists were still in the bank. The works affected both the Bank of America's own code and the software of third-party suppliers, so some of the problems found could not be corrected by one internal team.
Moynihan did not claim that all the vulnerabilities found only Mythos. It was about a new class of powerful models that can quickly find old defects in large code bases. Anthropic introduced the Claude Mythos Preview on April 7 and restricted access to the Model under Project Glasswing, where partners use AI to protect critical software analysis.
Anthropic stated that Mythos is able not only to find unknown mistakes, but also in some cases to build elements of exploits for vulnerabilities. The company reported thousands of potential problems of high and critical danger, but each finding before disclosure is checked by people. High search speed created a noticeable gap between automatic bug detection and manual correction check.
A similar effect has become noticeable outside the Bank of America. By early October, Claude had found about 23,000 potential problems, but people had managed to check less than 2000. Of the 1726, the tested 1726 were recognized as real. Massive vulnerability search is starting to scale markedly faster than triage, patch release, and secure implementation of updates.
The price of the works for Bank of America has not yet been disclosed separately. On the call with investors, Moynihan estimated the total cost of fixing the software and fulfilling the requirements against money laundering in a “couple of hundreds of millions of dollars with more” for the third quarter. The bank did not specify what proportion of the amount was on Mythos and AI-related findings.
In parallel, Bank of America faced large fines of regulators. FinCEN has imposed a $1.2 billion fine for violations of anti-money laundering requirements, with $750 million in OCC recovery counting in total. Regulators pointed to the shortcomings of transaction monitoring and the late submission of suspicious activity reports.
At the same time, AI remains a savings tool for the bank. Bank of America conducts about 130 internal projects with artificial intelligence and expects to get $250-300 million of the effect in 2026 and about $ 700 million by the end of 2027. Moinihan expects patching costs to come down after passing the main stream of problems found through the correction queue.
The contractors joined about 3000 employees who were already doing similar work. In mid-October, an additional thousand specialists were still in the bank. The works affected both the Bank of America's own code and the software of third-party suppliers, so some of the problems found could not be corrected by one internal team.
Moynihan did not claim that all the vulnerabilities found only Mythos. It was about a new class of powerful models that can quickly find old defects in large code bases. Anthropic introduced the Claude Mythos Preview on April 7 and restricted access to the Model under Project Glasswing, where partners use AI to protect critical software analysis.
Anthropic stated that Mythos is able not only to find unknown mistakes, but also in some cases to build elements of exploits for vulnerabilities. The company reported thousands of potential problems of high and critical danger, but each finding before disclosure is checked by people. High search speed created a noticeable gap between automatic bug detection and manual correction check.
A similar effect has become noticeable outside the Bank of America. By early October, Claude had found about 23,000 potential problems, but people had managed to check less than 2000. Of the 1726, the tested 1726 were recognized as real. Massive vulnerability search is starting to scale markedly faster than triage, patch release, and secure implementation of updates.
The price of the works for Bank of America has not yet been disclosed separately. On the call with investors, Moynihan estimated the total cost of fixing the software and fulfilling the requirements against money laundering in a “couple of hundreds of millions of dollars with more” for the third quarter. The bank did not specify what proportion of the amount was on Mythos and AI-related findings.
In parallel, Bank of America faced large fines of regulators. FinCEN has imposed a $1.2 billion fine for violations of anti-money laundering requirements, with $750 million in OCC recovery counting in total. Regulators pointed to the shortcomings of transaction monitoring and the late submission of suspicious activity reports.
At the same time, AI remains a savings tool for the bank. Bank of America conducts about 130 internal projects with artificial intelligence and expects to get $250-300 million of the effect in 2026 and about $ 700 million by the end of 2027. Moinihan expects patching costs to come down after passing the main stream of problems found through the correction queue.
