Artificial intelligence (AI) is transforming from a banking support tool to a potential threat that banks must be prepared to combat. Major US financial institutions are discreetly testing a new AI system that mimics cyber attackers. According to reports, officials associated with the Trump administration are advising banks to take this AI technology seriously and proactively assess how it could uncover vulnerabilities in their systems before malicious actors exploit them.
At the core of this advancement is Mythos, an innovative AI model created by Anthropic. Unlike conventional cybersecurity tools, Mythos is engineered to think like an attacker, identifying hidden weaknesses in systems and potential exploitation methods. This dual defensive and offensive capability has captured the interest of regulators.
Access to Mythos has been tightly controlled thus far, with JPMorgan Chase being one of the early adopters. However, other major players like Goldman Sachs, Citigroup, Bank of America, and Morgan Stanley are either testing or preparing to test the system internally, as per sources familiar with the situation.
The importance of this initiative was emphasized during a recent high-level meeting in Washington. Top Wall Street executives were summoned urgently by US Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell. While no specific threat was mentioned, officials cautioned banks to conduct stress tests against advanced AI capabilities before they become widely accessible.
This shift in mindset reflects policymakers’ growing concerns. While cybersecurity threats have always existed in banking, the advent of AI is amplifying the scale and speed of potential attacks. Mythos can not only identify individual vulnerabilities but also link multiple weaknesses to create significant breaches, a tactic historically challenging even for skilled human hackers.
Anthropic has underscored these risks through internal testing. For example, the AI successfully found ways to breach web browsers, enabling a malicious website to access data from other sites, potentially compromising sensitive financial information. Additionally, the model autonomously combined multiple vulnerabilities, a process that typically demands time, expertise, and coordination when done manually.
To mitigate these risks, the deployment of Mythos has been restricted to a select group of organizations under “Project Glasswing.” Apart from banks, companies like Amazon and Apple are also part of this initiative to secure critical systems and comprehend the model’s behavior before similar technologies become more widespread.
Government officials have stressed the urgency of this matter. Kevin Hassett recently stated, “It was appropriate that Secretary Bessent take the actions he did.” He further mentioned, “We are implementing every possible measure to ensure everyone is shielded from these potential risks, including Anthropic agreeing to delay the public release of the model until our officials have thoroughly evaluated it.”
</