The IMF Spring Meetings in Washington were supposed to be about monetary policy and debt crises. Instead, one topic dominated everything: Anthropic’s Mythos.
An AI model hijacks global financial policy
IMF Managing Director Kristalina Georgieva warned in an interview that the world lacks the ability “to protect the international monetary system against massive cyber risks.” Bankers, finance ministers, and regulators from around the globe weren’t debating interest rates — they were debating what happens when an AI model can find vulnerabilities in almost every computer on Earth.
And that’s not hyperbole. Anthropic’s Mythos has identified thousands of previously unknown zero-day vulnerabilities. Over 99 percent remain unpatched.
Bengio: ‘Private individuals are deciding the fate of everyone else’
Yoshua Bengio — one of the three “Godfathers of AI” and Turing Award winner — used the occasion to deliver a pointed warning. Speaking to Fortune, he said: “It doesn’t make sense that private individuals are deciding the fate of infrastructure for everyone else.”
His argument is simple and uncomfortable: Anthropic alone decided who gets access to Mythos. Apple, Google, Microsoft, Nvidia, JPMorgan Chase — they got in through “Project Glasswing.” Everyone else? They wait and hope their systems don’t get hacked in the meantime.
Bengio is calling for a regulatory body modeled on the FDA — an institution that oversees the deployment of advanced AI systems. And he goes further: international agreements are needed, including with China.
What this means
You don’t have to fully agree with Bengio to see the problem. A single company owns a tool that affects the cybersecurity of the entire global infrastructure — and unilaterally decides who benefits from it.
European regulators and bank executives at the IMF meetings openly admitted they have no access to Mythos and cannot assess the scope of the vulnerabilities it has uncovered. That’s not a great look for a technology that’s supposed to help the whole world.
The question is no longer whether AI needs regulation. The question is whether regulation can arrive fast enough — or whether the next zero-day will be faster.
Sources: