Bank of America has engaged 1,000 contractors to address software vulnerabilities identified by the Claude Mythos AI model. Developed by Anthropic and released in early April 2026, the model is designed for cybersecurity tasks, including the autonomous discovery of code flaws.
The model was provided to select US banks through the Project Glasswing initiative. During testing, one scan identified over 23,000 potential issues. While most findings were of low to moderate severity, the process also revealed critical vulnerabilities within internal systems.
Bank of America began testing the model by mid-2026. The implementation of Mythos reduced the time required to address security gaps from weeks to days. The bank is currently expanding its operational capacity to manage the remediation process.
JPMorgan CEO Jamie Dimon noted that the use of such AI tools could increase cyber risk exposure by a factor of ten. Consequently, the US Treasury and the Federal Reserve have conducted briefings for bank executives regarding these risks and the security of legacy systems.
As of October 2026, access to Mythos remains restricted. The model is available only to vetted cybersecurity and research professionals through controlled programs and has not been released for public use.
