To all Banks This Is Something You Need To Watch

Anthropic’s Claude Mythos Preview was released this week in limited capacity, deliberately constrained because the company itself acknowledged that hackers could exploit its capabilities. Let that sink in. A frontier AI lab launched a model it openly concedes could be weaponized, briefed senior U.S. government officials on its offensive cyber applications, and simultaneously rolled it out to JPMorgan, Apple, Google, Microsoft, and Nvidia under something called Project Glasswing. The Fed Chair and Treasury Secretary then convened an emergency session with the CEOs of Bank of America, Citigroup, Goldman Sachs, Morgan Stanley, and Wells Fargo to discuss the threat.

This is the new normal. And most financial institutions are nowhere near ready for it.

Here is what bankers need to understand right now. Anthropic’s own models were used last November by a Chinese threat group to automate attacks on government and corporate targets. The company disclosed it. Quietly. And the market kept moving. Meanwhile, a draft blog post leaked to Fortune revealing Mythos’s advanced capabilities sent cyber stocks into a slump, not because the technology was surprising to insiders, but because the public got an unfiltered look at what these models can actually do in adversarial hands.

The Department of Defense has labeled Anthropic a supply chain risk to national security. A federal appeals court this week denied the company’s bid to block that designation. The U.S. government is simultaneously treating Anthropic as a strategic partner and a potential liability (really depends on the day). Both things are true. That ambiguity is precisely what makes this moment so dangerous for the financial sector.

Your third-party and fourth-party risk frameworks were not built for this. Your vendor management programs assume a relatively stable threat surface. AI changes the attack velocity, the attack creativity, and the attack scale, simultaneously. What used to require a sophisticated nation-state actor can now be automated, iterated, and deployed at low cost. The barrier to entry for a highly capable cyberattack just dropped again.

The banks in that room with Powell and Bessent are the ones with the resources and the relationships to get early warning. Most of your institutions are not in that room. They are reading about it the next morning, the same as everyone else.

The question is not whether AI-powered threats will hit the financial sector. They already have. The question is whether your risk program is calibrated for the speed and sophistication of what is coming, or whether you are still fighting the last war with last decade’s controls.

Get ahead of this. The cost of being wrong has never been higher.