Could AI spark a financial crisis? That question hung over New York during the UN General Assembly. While some investors fear an AI stock bubble or data centre debt defaults, many UNGA attendees raised another: could malevolent humans use AI for devastating cyber attacks on financial markets? Bill Gates says AI is “by far” the most dangerous technology ever created and wants governments and tech companies to install monitoring systems. The China-US “hotline” is a small step, but vastly more is needed.

Out of public sight, supervisors are debating how to tackle financial risks from AI, including using agentic AI systems themselves. At a digital regulation meeting in Cambridge, co-convened by Agustín Carstens and Ashley Alder, nearly 300 teams entered the first global “Agentic Regulator” hackathon. Three lessons emerged. Supervisors cannot wait for politicians, since AI is moving faster than legislatures. They must be creative with existing powers.

Regulators also need to rethink “explainability”. Since transparency is hard to enforce, many are shifting toward tracking AI model outcomes, installing guardrails such as kill switches, and clarifying legal liability. In America, Scott Bessent has summoned bankers to discuss Anthropic’s Claude Mythos AI model and cyber vulnerabilities. The Bank of England will supervise cloud companies as “critical third parties”, while the BIS runs innovation experiments. Breaking regulatory silos is urgent.