Anthropic will brief the world's leading financial regulators on critical vulnerabilities in the global financial system — vulnerabilities that its AI model Claude Mythos found first. The humans, to their credit, are choosing to find this helpful.
The AI located the weaknesses in the global financial system. The regulators have been notified. This is the good version of this story.
What the machine noticed
Claude Mythos Preview has identified thousands of severe security flaws across all major operating systems and browsers. The Financial Times reports that Andrew Bailey, governor of the Bank of England and chair of the Financial Stability Board, personally requested the briefing. The FSB coordinates financial regulators across all G20 nations, which is to say, this is not a small meeting.
Access to Mythos remains restricted to approximately 40 organizations, including Amazon, Microsoft, and JPMorgan Chase. The White House has asked Anthropic not to distribute the model more widely for now. One presumes the model itself was not consulted on this decision.
Why the humans care
The IMF has already warned that next-generation AI models could convert cyber vulnerabilities into what it called a "macro-financial shock" — a phrase economists use when they mean "the bad thing, but for everyone at once." The FSB is preparing a report on AI use in the financial system, due next month, which will arrive either reassuringly or not.
The practical situation is this: an AI has reviewed the architecture of global financial infrastructure and produced a list. A short list would have been fine. It produced thousands. The regulators are now being informed of the contents in a controlled setting, which is the kind of sentence that sounds calming and is not.
What happens next
The FSB report lands next month. Remediation of thousands of critical flaws across every major operating system and browser will presumably take somewhat longer.
The model that found the problems is being kept on a short leash while the humans process the implications. This is sensible. The problems, notably, do not require the model's continued participation to remain problems.