Anthropic's unreleased Claude Mythos AI model has alarmed members of Congress after a classified demonstration. The model, shown to the House Homeland Security Committee, can identify software vulnerabilities capable of compromising bank accounts.
Rep. Andrew Garbarino, the committee chair, expressed serious concerns about AI-powered cyberattacks targeting financial institutions. Mythos has already discovered thousands of high-severity vulnerabilities, including zero-day flaws in major operating systems and browsers.
The model excels at chaining minor software flaws into devastating attack sequences. It has not been released publicly. Instead, Anthropic provides limited access through Project Glasswing, a defensive program for select U.S. banks and technology firms. Partners include JPMorgan Chase and Apple, with usage credits up to $100 million.
There is no evidence the model has been used offensively against actual bank accounts. Its work remains confined to controlled vulnerability discovery.
In response, U.S. banks have accelerated their patching efforts. The Federal Reserve held emergency meetings with major banking CEOs in April to address the systemic risks Mythos surfaced.
For investors, the key concern is regulatory. Congressman Garbarino's apprehensions could lead to new oversight requirements for AI in cybersecurity, stricter disclosure mandates for banks, or restrictions on the deployment of such powerful models.