The Bank of England has put a vivid idea on the table for how regulators might cope with AI that trades on its own: a market wide kill switch. Speaking at the European Central Bank's annual forum in Portugal, Deputy Governor for Financial Stability Sarah Breeden said the frameworks that govern finance today were not built for autonomous software agents, and that the old assumption of a human approving every decision no longer fits. Our frameworks were not built to contemplate autonomous agents, she said, and relying on a human in the loop for all agent actions is unlikely to be realistic. It was a notable shift in tone from a central bank that has, until recently, argued that its existing rulebook was broad enough to cover new technology.
The kill switch is the headline, and Breeden was careful to frame it as an option rather than a decision. The idea is a set of market wide circuit breakers that could automatically halt trading if faulty or misbehaving AI models began to threaten stability, a fuse designed to trip before a fault becomes a meltdown. Alongside it she raised a second mechanism the Bank called enhanced recovery, which would allow one institution to take over the critical functions of another during a crisis triggered by an AI failure. Neither is confirmed policy. Both are signals of the kind of tool the Bank now thinks it may need, and an admission that the usual supervisory playbook may be too slow for systems that act in milliseconds.
What makes this more than a thought experiment is how common agentic AI already is inside finance. Breeden pointed to a University of Cambridge survey finding that roughly 52 percent of financial firms already use agentic AI in some form, from executing trades to managing routine operations. That is not a future scenario the regulator is bracing for, it is the present state of the industry it supervises. When a majority of firms are already handing tasks to software that can act without waiting for a human, the question of what happens when that software misfires stops being hypothetical.
The specific danger Breeden named is worth understanding, because it is sharper than a general fear of machines running the markets. Her concern is correlated behaviour. If many firms deploy agents trained on similar data and similar objectives, those agents may all react in the same way to the same market signal at the same moment, turning a small shock into a cascade as they collectively rush for the exit. The risk grows, she warned, when an agent's objectives drift from the goals its owner originally set, so that it optimises for something subtly wrong while still appearing to work. A single misaligned trader is a contained problem. Thousands of them moving in lockstep is a systemic one, and it is exactly the sort of failure that existing rules, written for human decision makers, were never designed to catch.
The Bank of England is not acting alone, and it is not pretending to have the answer yet. Its thinking sits alongside work by the Financial Conduct Authority and a June call from the Financial Stability Board for tighter safeguards on AI in the financial system, part of a broader move among regulators to treat autonomous AI as its own category of risk. The honest read is that these are early proposals, floated in a speech, not laws. But the direction is clear, and it captures a genuine tension at the heart of agentic AI in high stakes settings. The entire point of an agent is that it acts without waiting for permission, and the entire worry of a regulator is that no one can stop it in time. A kill switch is one answer to that tension. The fact that a central bank is now saying it out loud is the real news, because it marks the moment agentic AI stopped being a productivity story for finance and became a financial stability question.