Bank of America’s Mythos Patch Sprint 2026: AI Bug Hunting in Banking

Bank of America has reportedly hired about 1,000 contractors to fix software flaws found by Anthropic’s Mythos model. That one staffing decision says more about AI vulnerability scanning in banking than a year of vendor keynotes. When a top U.S. bank brings in a contractor army to patch bugs a machine found, AI bug hunting has moved past the pilot stage. It is now a budget line. Finding flaws is no longer the slow part. Fixing them is.

What’s actually new: AI vulnerability scanning in banking hits production scale

According to reports, Bank of America is running a large remediation sprint with roughly 1,000 outside contractors. They are working through software weaknesses surfaced by Anthropic Mythos, a frontier model that Anthropic has limited to a small group of partners because of its offensive security capabilities. Neither the bank nor Anthropic has published a full breakdown of how many flaws were found, how severe they are, or which systems are affected, so treat the specifics as reported, not confirmed. The direction is clear either way: AI-discovered software flaws are arriving faster than in-house teams can close them.

The difference is volume and age. Earlier automated scanners mostly flagged known vulnerability patterns and outdated dependencies. Anthropic has said Mythos-class models can reason through large, old codebases and find exploitable bugs that went unnoticed for years, sometimes decades. Banks run decades of layered code: COBOL cores, Java middleware, mobile apps, and third-party integrations. For them, a backlog of hidden technical debt becomes a visible to-do list all at once.

Visibility creates obligation. Once a bank knows about a flaw, regulators, auditors, and eventually plaintiffs’ lawyers expect it to act. That explains why the reported response is a staffing surge and not a slide deck. Legacy system remediation is no longer a someday project. It is a sprint with deadlines.

Why it matters

• Finding bugs is getting cheap. Fixing them is not. AI moves the security bottleneck from discovery to remediation. Budget for engineers, testing, and change management, not just scanning tools.
• Attackers get the same capability. Restricted access buys defenders time, but similar models will spread. Any flaw an AI can find for a bank, a well-funded attacker’s AI can eventually find too. The window between “discoverable” and “exploited” is shrinking.
• Compliance exposure goes up. Under frameworks like the NYDFS cybersecurity regulation, PCI DSS 4.0, and the EU’s DORA, a known, documented, unpatched vulnerability is far worse than an unknown one. AI scanning creates paper trails fast.
• Vendors and suppliers are next. Banks will push these expectations down the supply chain. If you sell software or services to a financial institution, expect tougher security questionnaires, requests for proof of patch SLAs, and possibly requests to scan your code.
• The labor market moves. Bank patch management in 2026 needs people who can read old code, write regression tests, and ship safe fixes. That helps contract developers and managed security providers and raises costs for everyone else.
• Small firms aren’t exempt. Community banks, credit unions, fintechs, and the businesses that integrate with them all rely on the same open-source libraries. A flaw in shared code affects everyone downstream.

How to use AI vulnerability scanning today

You probably can’t get Mythos, because access is limited. You can still run the same workflow (find, triage, fix, verify) with free or nearly free tools. This playbook is for a business owner with a small dev team or an outside agency.

1. Inventory what you actually run. You can’t patch what you don’t know about. Generate a software bill of materials (SBOM) for each application.

   # Install Syft and generate an SBOM for a project folder
   curl -sSfL https://raw.githubusercontent.com/anchore/syft/main/install.sh | sh -s -- -b /usr/local/bin
   syft dir:./my-app -o cyclonedx-json > sbom.cdx.json

2. Scan dependencies for known flaws. Most real-world breaches still come through outdated libraries. Google’s OSV-Scanner is free and fast.

   # Scan a repository against the OSV vulnerability database
   osv-scanner scan source -r ./my-app
   
   # Or scan the SBOM you just generated
   osv-scanner scan source -L sbom.cdx.json

3. Run static analysis on your own code. Semgrep’s community rules catch injection, auth mistakes, and hardcoded secrets in most common languages.

   pip install semgrep
   semgrep scan --config auto --severity ERROR --json -o semgrep-results.json ./my-app

4. Add an AI reviewer for logic flaws. Pattern scanners miss business-logic bugs, such as a refund endpoint that doesn’t check ownership. A general-purpose AI coding assistant can catch these. If your team uses Claude Code, it includes a built-in security review command:

   # Inside a repository, from the Claude Code CLI
   /security-review

   With any capable AI assistant, use a focused prompt like this one:

   You are a senior application security engineer reviewing code for a
   business that handles payments and customer financial data.
   
   Review the attached files and report ONLY exploitable issues.
   For each finding, give:
   1. File and line number
   2. Vulnerability class (e.g., IDOR, SQL injection, broken auth)
   3. A concrete exploit scenario in 2-3 sentences
   4. Severity (Critical/High/Medium/Low) with a one-line justification
   5. A minimal code fix
   
   Do not report style issues or theoretical concerns without an exploit path.
   If you are not confident, say so and explain what you would need to verify.

5. Triage like a bank: severity times exposure. A thousand findings are useless without a ranking. Score each one by how severe it is and how reachable it is (internet-facing, handles money or PII). Fix critical, internet-facing issues first.

   Triage rubric (copy into your tracker)
   - P0: Critical/High + internet-facing + touches payments or PII -> fix in 72 hours
   - P1: Critical/High + internal only, or Medium + internet-facing -> fix in 14 days
   - P2: Medium + internal -> fix in 30 days
   - P3: Low -> next scheduled release
   Every item needs: owner, due date, fix PR link, verification note.

6. Automate the boring fixes. Turn on automated dependency updates so routine patches start without a human.

   # .github/dependabot.yml
   version: 2
   updates:
     - package-ecosystem: "npm"
       directory: "/"
       schedule:
         interval: "weekly"
       open-pull-requests-limit: 10
       groups:
         security:
           applies-to: security-updates
           patterns: ["*"]

7. Verify, then document. Re-run your scans after every fix and keep the before-and-after reports. If a bank partner, insurer, or regulator asks, that evidence trail protects you.

8. Ask your vendors the hard question. Your risk includes their code. Send something like this to every software provider that touches your money or customer data:

   Subject: Security remediation practices - request for confirmation
   
   As part of our 2026 vendor review, please confirm:
   1. Whether you use automated or AI-assisted vulnerability scanning, and how often
   2. Your patch SLAs for critical and high-severity findings
   3. Whether you maintain an SBOM for the product we use, and can share it
   4. Any critical vulnerabilities disclosed in the last 12 months and their fix dates
   
   Please reply within 15 business days.

How it compares

Mythos isn’t the only AI bug hunter. Here’s how the main options compare for financial services AI security. Details on restricted tools come from public announcements and may change.

Tool	Who it’s for	Strength	Availability
Anthropic Mythos	Large enterprises, critical infrastructure partners	Deep reasoning over large, legacy codebases; finds novel, long-hidden flaws	Restricted partner access
Google Big Sleep / CodeMender	Open-source and Google ecosystem projects	AI-driven discovery plus automated patch proposals	Research and limited programs
OpenAI Aardvark	Development teams wanting an autonomous security agent	Continuous repo monitoring with exploit validation	Private beta
GitHub Copilot Autofix (CodeQL)	Teams already on GitHub	Tight CI integration, suggested fixes on pull requests	Generally available; free for public repos
Semgrep / OSV-Scanner	Small businesses and lean teams	Fast, cheap, well-understood rule-based scanning	Free community editions

Our take: most businesses don’t need frontier-model access to get 80% of the value. They need consistent scanning, a triage process, and the discipline to ship fixes. Frontier tools matter most for organizations with huge legacy codebases, which is exactly why a bank went first.

What’s next

Expect other large banks to launch their own remediation programs, either quietly through contractors and consultancies or publicly as part of resilience reporting. Once one major institution shows what AI-assisted scanning turns up, its peers will struggle to explain to their boards why they haven’t looked. Watch for large consulting and outsourcing firms to sell “AI-found flaw remediation” as a standalone service with fixed per-finding pricing.

Regulators come next. Supervisors already ask how banks govern AI models. Their next question is how banks govern what those models find. Guidance could cover expected remediation timelines for AI-discovered vulnerabilities, disclosure duties when flaws affect shared vendors, and whether skipping available scanning tools counts as a control gap. Firms covered by DORA and NYDFS should assume examiners will ask.

Finally, watch access widen. Anthropic and its competitors have signaled that restricted security models are a bridge, not a permanent state. As defensive access expands and comparable capabilities appear elsewhere, the advantage shifts to whoever fixes fastest. The winners won’t be the businesses with the best scanner. They’ll be the ones that can patch, test, and deploy in days instead of quarters.

Frequently Asked Questions

What is Anthropic Mythos?

Mythos is a frontier AI model from Anthropic with unusually strong code analysis and security research abilities, including finding previously unknown vulnerabilities. Because those abilities could be misused, Anthropic has limited access to selected partners focused on defense.

Did Bank of America confirm the 1,000-contractor figure?

No. The number comes from news reports, not a detailed public disclosure from the bank, so treat the exact headcount as reported. The broader trend is well supported: large institutions are scaling up remediation to handle AI-discovered software flaws.

Can a small business use AI vulnerability scanning without a big budget?

Yes. Free tools like OSV-Scanner, Semgrep community rules, Dependabot, and GitHub’s CodeQL cover most common risks. Add an AI coding assistant to review business logic, then focus your effort on a clear triage and fix process.

Why do banks struggle so much with patching?

Legacy banking systems often run decades-old code that few current employees fully understand. Every change needs testing, approvals, and controlled releases to avoid outages, so a fix that takes an hour to write can take weeks to deploy safely. That makes remediation, not discovery, the bottleneck.

Does this affect businesses that only partner with banks?

Very likely. Banks pass security expectations on to vendors, fintech partners, and service providers. Expect more detailed questionnaires, requests for SBOMs, and contract clauses that set patch timelines.

Will AI make banks more secure or less?

In the short term, both. Defenders with early access can close old holes before attackers find them, but the same capability will eventually reach attackers. Banks that build fast remediation pipelines now will come out ahead. Banks that only buy scanners will end up with longer to-do lists.