Speaking to a room of central bankers in Sintra, Portugal, one of the Bank of England's most senior officials floated an idea that would have sounded like science fiction a few years ago: a switch to turn off the machines. Sarah Breeden, the Bank's deputy governor for financial stability, told the European Central Bank's annual forum that regulators may need "circuit breakers or kill switches that would limit or stop trading market-wide if faulty AI models cause market meltdown." The warning, delivered on June 30 and reported by Reuters, marks the clearest signal yet that Britain's financial watchdog is preparing for a future in which autonomous software, not humans, drives markets at machine speed.
Breeden's concern is not that any single AI trader misbehaves. It is that many of them behave the same way at the same moment. She said Bank officials have been running simulations of what would happen if AI-powered traders all tried to execute similar trades simultaneously, and that the Bank has been researching how to combat so-called "herding behaviour" by AI tools alongside Germany's Bundesbank and the Bank for International Settlements. The capabilities of AI, she warned, "could amplify volatility" when markets come under stress.
A shift in tone from Threadneedle Street
What makes the intervention notable is how much it departs from the Bank's previous posture. For years, the official line was that existing, technology-neutral rules were sufficient to supervise whatever tools firms chose to deploy. Breeden signaled that this confidence is fraying.
"Our frameworks were not built to contemplate autonomous agents, and relying on a human in the loop for all agent actions is unlikely to be realistic," she said, according to reporting on her remarks. Her worry centers on correlated behavior: if many agents respond identically to the same trigger, they could amplify a shock, "especially if their objectives drift from original goals."
She was equally blunt about the pace of change. Breeden noted that the capabilities of AI systems doubled roughly every seven months in 2019, a timeframe that had shortened to around four months by 2024. "An already exponential increase in capability appears to be accelerating," she said. She added a challenge aimed squarely at her fellow regulators: "As AI capabilities increase, we must keep asking whether existing, technology-agnostic regulatory frameworks remain sufficient."
From algorithms to agents
The Bank has been here before, in a smaller way. Breeden pointed to 2016, when it investigated a "flash crash" in the pound that it concluded was probably caused by a cluster of computer algorithms executing trades simultaneously during a quiet part of the trading day. Algorithmic trading is old news; growing numbers of transactions already run on code that executes almost instantly.
What is new is the leap from algorithms to "agentic" AI, systems that can set goals, plan, and take actions without constant human supervision. The Bank's simulations are now examining what happens when such agents are wired into trading desks. For now, Breeden said, trading firms mostly confine autonomous AI to lower-risk operational tasks such as research, but she warned that this could change quickly, with consequences for markets.
The numbers suggest the shift is already underway. Breeden cited a Cambridge survey finding that 52 percent of finance firms already use agentic AI. The Bank and Financial Conduct Authority's own survey of AI in UK financial services found that 55 percent of respondents' AI use cases involve some form of autonomous decision-making, though only about 2 percent are described as fully autonomous today. The gap between those figures is precisely the space regulators are trying to get ahead of.
Why this is a market-structure problem
The kill-switch framing captures attention, but the deeper issue is one of market structure. Traditional circuit breakers, the exchange-level halts that pause trading after sharp moves, were designed for a world of human panic; they assume that stopping the clock gives people time to reassess and step back from a stampede. The relevance is not theoretical: South Korea's Kospi index triggered its circuit breaker twice in a single week during recent volatility, having used it only 11 times since 2000.
AI herding scrambles that logic. If dozens of models are trained on similar data, optimize for similar objectives, and react to the same signals in microseconds, a market can move further and faster than any human-paced backstop was built to contain. Breeden's point about objectives that "drift from original goals" hints at a subtler danger: agents that pursue strategies their operators no longer fully understand, then reinforce one another's behavior in a feedback loop no single firm controls.
That is why the mitigations under discussion go beyond a simple off switch. Breeden also floated "enhanced recovery" measures that would let one bank take over another's core functions in a crisis, alongside severity thresholds and recovery obligations. The Financial Stability Board made a similar call in June for tighter safeguards against agentic AI, and the FCA has been openly rethinking its approach for the AI age.
The intervention lands amid broader anxiety about AI in finance. Days earlier, the Bank for International Settlements warned that "excessive" spending on AI data centers and opaque financial ties between AI giants, shadow banks, and data-center builders risked a meltdown reminiscent of the 2008 credit crunch. Trading behavior is only one of several AI channels regulators now see running toward systemic risk.
What to watch
Breeden framed kill switches as options rather than firm policy, but as one analysis of her speech put it, the Bank rarely trails ideas it has no intention of pursuing. The direction of travel is clear. Watch the Bank's simulation results, which will shape whether "herding" becomes a formal category in future stress tests. Watch the Financial Policy Committee's next financial stability assessments for language on agentic AI and market-wide safeguards. Watch, too, for coordination, or the lack of it, between the Bank, the FCA, the Bundesbank, and the BIS, because a kill switch that halts trading in London but nowhere else offers little protection in a globally wired market. And watch the practical question underneath all of it: whether any switch, however well designed, can act fast enough to matter once the machines are already moving.
"Our frameworks were not built to contemplate autonomous agents, and relying on a human in the loop for all agent actions is unlikely to be realistic."— Sarah Breeden, Deputy Governor for Financial Stability, Bank of England