John Z Black
Brunswick, ME • (207) 245-1010 • contact@johnzblack.com
On April 10, Treasury Secretary Scott Bessent and Fed Chair Jerome Powell sat down with America’s largest bank CEOs at Treasury headquarters and said: this AI model is a problem you need to take seriously right now.
The joint appearance of those two at a single private-sector meeting is not a routine event. Both organizations confirmed follow-on meetings were planned. White House NEC Director Kevin Hassett publicly confirmed it on Fox News. This happened.
The specific risk argument comes down to how banking infrastructure is built. Every major bank uses roughly the same vendors for KYC, onboarding, and transaction processing. Regulatory requirements push institutions toward a narrow set of compliant providers. The result is a shared infrastructure surface that’s nearly uniform across the sector. Mythos, per Anthropic’s own characterization, can identify and exploit weaknesses across every major operating system and browser. Against infrastructure that looks almost identical at every institution, a uniform-failure scenario isn’t theoretical.
Here’s the part that got less attention. The government didn’t just warn banks. Major banks were quietly given private access to Mythos through Project Glasswing. JPMorgan, Amazon, Apple, and several dozen others. The framing is defensive: let defenders study the capability before adversaries get there. That logic isn’t wrong.
But the governance mechanism for a designated systemic risk is that one company, Anthropic, decides who gets access and when. When something gets classified as a threat to financial infrastructure, the normal response involves some kind of regulatory framework. What actually happened is that the government convened an emergency meeting, then deferred to the vendor on access control.
Canada ran a parallel session. Their Finance Ministry and Bank of Canada held similar discussions with Canadian bank executives in roughly the same window. Two countries, same week, same concern.
Some industry voices think the coverage has been overheated. IBM researcher Lav Varshney: “I don’t think we’ve entered some new era.” That’s a fair counterpoint. The gap between demonstrated capabilities and worst-case projections is real.
But Bessent and Powell apparently made their assessment. The meeting is documented. The question of what happens as more institutions want Glasswing access, or as adversaries develop comparable capabilities independently, is one regulators haven’t answered yet.