Bank of America adds 1,000 contractors to fix software flaws exposed by AI

Bank of America added 1,000 contractors to repair software flaws exposed by advanced AI models, CEO Brian Moynihan disclosed October 14. They remain alongside 3,000 internal staff. The disclosure puts a human workload behind the promise of AI-assisted security: Anthropic says its Mythos model, among those involved, was released to help secure critical software.

The repair work reaches vendor code

Speaking to CNBC’s Leslie Picker after the bank’s third-quarter earnings release, Moynihan said the contractors joined in the second quarter. The work covers both the bank’s own software and code supplied by vendors. Anthropic’s Mythos is one of the advanced models he identified as exposing vulnerabilities that required repairs.

The distinction between outside help and existing staff matters. Moynihan described the contractors as an addition to 3,000 people already working in that area, not 4,000 new hires. Months after the second-quarter hiring, the outside workers were still helping the bank patch its systems.

Moynihan called the risk from frontier models real and described an urgent effort to work through it.

A large bill, but no Mythos-only total

Moynihan told analysts the bank spent a “couple hundred million plus” in the third quarter on software remediation and regulatory issues involving anti-money-laundering compliance. He gave no separate vulnerability-patching figure, as The Decoder noted in its coverage of the CNBC interview. The combined expense is not a standalone bill for Mythos-related repairs.

Moynihan expects patching costs to decline over time as the bank works through the repairs. His explanation was that completing patches reduces the amount of patching left to do. That is management’s expectation for the workload, rather than a disclosed timetable for finishing it.

Why finding a flaw is only the start

Anthropic’s April 7 technical assessment explains the defensive rationale behind Mythos Preview. The company initially limited access through Project Glasswing to critical industry partners and open-source developers. Its stated aim was to let defenders secure important systems before models with similar capabilities became broadly available.

The company said the same improvements in coding, reasoning and autonomy made Mythos better at both patching flaws and exploiting them. Its researchers described finding previously unknown vulnerabilities, including bugs that had survived for decades. Those are Anthropic’s findings about its model, not an independent assessment of Bank of America’s systems.

Anthropic’s research process also separates generating a finding from confirming it. In its described tests, an agent examined code, experimented with suspected weaknesses and produced a bug report with reproduction steps when successful. A further model pass checked whether the finding was real and important.

Human review followed: Anthropic said it sent the highest-severity bugs to professional reviewers for validation before notifying software maintainers. At publication in April, fewer than 1% of the potential vulnerabilities it had discovered were fully patched. The company said that review process was intended to avoid overwhelming maintainers, while acknowledging that it lengthened the path to a completed fix.