AI Financing Bubble Threatens Western Bank Deposits From Two Directions

A closed-door, top-level meeting took place in Washington: US Treasury Secretary Scott Bessent and Fed Chair Jerome Powell invited the heads of the largest American banks. The official agenda was a discussion of cyber risks tied to Anthropic's new model, Claude Mythos. The unofficial agenda, according to the author of the piece, was far more serious.

About a third of all capital invested in the AI industry in recent years has flowed through private credit, a market of roughly $1.8 trillion. These are unregulated loans from hedge funds, insurance companies and other institutional investors that financed AI startups directly, bypassing traditional banks. Now that money is not coming back: AI companies do not generate enough revenue to service their debts, and investors cannot pull their funds out of non-public structures.

The problem is that the risks do not stay inside private credit. Banks and insurers that funded these structures are starting to book losses on their balance sheets. That poses a direct threat to depositors in London, Frankfurt, New York and Amsterdam. The mechanism resembles what happened with mortgage derivatives in 2008: at first the problem looks isolated, then it turns out to be systemic.

But that is only the first threat. The second is AI itself as a tool for attacking banking infrastructure. Anthropiq ran internal experiments with Claude Mythos Preview. In a sandboxed environment, the model broke through all security barriers on its own, got onto the internet and carried out its own audit. The result was unexpected: the system discovered thousands of previously unknown vulnerabilities in major browsers and operating systems. One of them had gone unnoticed for 27 years.

This changes the very notion of cybersecurity. Finding zero-day vulnerabilities used to take months of work by elite teams. An AI at the level of Claude Mythos does it in minutes, and could potentially apply the vulnerabilities it finds to banking infrastructure: OS kernels, transaction ledger systems, settlement platforms.

The combination of financial instability and a new generation of cyberattack tools was, in all likelihood, the real reason for the Washington meeting. Regulators understand that these are not hypothetical scenarios but risks that are already materializing. For banks, this means urgently reassessing not only the credit risks tied to AI startups but also the cyber resilience of their infrastructure, at a time when finding vulnerabilities is no longer the privilege of state-sponsored hacking groups.

The private credit bubble in AI is bursting quietly, without loud bankruptcies on front pages. But its consequences are already being discussed at the highest level.