Kill switches for the stock market: Inside the Bank of England’s AI contingency planning

The Bank of England is openly discussing whether financial markets will need system-wide kill switches to shut down trading if autonomous AI agents send them into meltdown.

Speaking at the European Central Bank’s Sintra Forum on 30 June, Deputy Governor for Financial Stability Sarah Breeden said the financial system was likely to evolve quickly into one that “operates more autonomously, at scale and speed”, with AI agents devising and executing trading strategies rather than simply assisting the humans who do.

For now, the evidence suggests trading firms mostly deploy autonomous AI for lower-risk operational work such as research. But Breeden warned that could change quickly, and that the consequences of getting the transition wrong could be systemic.

The herding problem

The core worry is not that a single AI trader goes rogue, but that many of them behave the same way at the same time.

If AI agents respond similarly to the same prompts or market triggers, Breeden said, they could amplify volatility in a stress event. This is more likely if their objectives drift from what their operators intended, a manifestation of the “misalignment” problem that AI safety researchers have documented in some models.

This changes the dynamic from whether individual firms can use AI models responsibly and is instead a question of whether the system as a whole can observe and contain the collective behaviour those models produce. A thousand well-governed trading agents can still stampede together.

To combat this, the Bank of England has run structured, multi-round stress simulations that examine how market participants behave as they learn of each other’s actions. Breeden said AI made that work more important, and should transform how it is done.

The Central Bank is now experimenting with the Bank for International Settlements’ Innovation Hub and Germany’s Bundesbank on simulation methods designed to identify which aspects of agent design could drive herding behaviour.

Circuit breakers for the machine learning age

Breeden said that work would likely explore whether markets using AI agents are resilient enough, whether the agents’ objective functions could be made to incorporate public policy goals, and whether guardrails are needed “analogous to circuit breakers or kill switches that would limit or stop trading market-wide if faulty AI models cause market meltdown”.

Circuit breakers already exist on major exchanges as automatic trading halts triggered by sharp price falls, introduced after the 1987 crash and refined after the 2010 flash crash.

But these mechanisms are designed for a world where humans, however panicked, are ultimately making the decisions. A market populated by autonomous agents operating at machine speed may need something more drastic: the ability to switch the machines off entirely.

Breeden went further still, floating the idea that agents’ objective functions, the goals an AI system is built to pursue, could themselves incorporate public policy objectives.

That would mark a remarkable intervention as regulators reach inside the design of private trading systems rather than policing their outputs after the fact.

Breeden acknowledged this and noted that existing regulatory frameworks are technology-agnostic and were not built to contemplate autonomous agents. She added that relying on a human in the loop for every agent action is unlikely to be realistic and that more sophisticated governance and accountability frameworks may be needed.

A matter of urgency

The speech landed a week before the Bank’s Financial Policy Committee publishes its updated assessment of AI and financial stability on 7 July, and Breeden was careful not to front-run it. But it’s clear what direction the Bank of England is heading in.

AI capability, she noted, is on an accelerating exponential. The length of software task that leading models could complete was doubling roughly every seven months in 2019; by 2024 that had compressed to around four months, and advances in models for identifying cyber vulnerabilities suggest it may now be faster still.

“We were surprised this Spring,” she said, “and we should be prepared for further technology surprises.”

The Bank is also exploring whether agentic AI could transform its own surveillance of markets. This would involve combining qualitative and quantitative data to generate risk scenarios, potentially using a “digital twin” of the financial system to simulate how participants interact under stress.

What’s clear from he speech is that nobody, including the Bank, knows how a market full of autonomous traders will behave in a crisis.

Talks of building a kill switch before it is needed, rather than designing it in the wreckage afterwards, is about as clear a statement of intent as a central banker gives.

Now read: Andy Burnham thinks the UK’s big tech romance is over: report