On June 30, 2026, Bank of England Deputy Governor Sarah Breeden stood at the ECB's annual forum in Sintra, Portugal, and proposed something that already exists. She called it a market-wide kill switch: a mechanism that would automatically suspend trading if AI agents triggered predefined instability indicators. She warned that AI agents trained on similar data could exhibit herd behavior, making identical trading decisions simultaneously. She described it as the most significant intervention on AI trading risk from any central bank.
Three days later, Meta announced it would sell excess AI compute at a discount, and KOSPI triggered a sidecar halt as Samsung and SK Hynix fell 9 and 15 percent respectively. It was the fifth market intervention on the Korean exchange in 2026. Circuit breakers had already fired four times: on March 4 and March 9 during the Iran strikes, on June 8 during a semiconductor rotation, and on June 23 during another AI tech selloff. Every trigger was a human decision.
The March circuit breakers fired when US-Israeli strikes on Iran sent oil past $120 and the Strait of Hormuz closed. June 8 was a global semiconductor rotation that sent the index down 8.3 percent. June 23 was another wave of AI chip selling. July 2 was a corporate announcement from Menlo Park. Military decisions, sector rotation, press releases. No AI agent made any of those trades. The herd behavior Breeden wants a kill switch to prevent is already happening without AI, and the existing kill switch is already catching it.
Breeden's speech came with a number. A Cambridge Centre for Alternative Finance survey, conducted in partnership with the BIS, IMF, and World Economic Forum, found that 52 percent of finance firms already use agentic AI. Twenty-three percent are at mature adoption stages. Eighty-one percent of respondents say agentic AI will be meaningfully achieved by 2030. The regulator proposed to build the infrastructure after a majority of the industry had already adopted the technology.
This is not the first time financial regulators arrived late with the right idea. On August 1, 2012, Knight Capital Group lost $440 million in 45 minutes because it lacked a kill switch. A software deployment error activated dormant trading code that executed millions of erroneous trades across 148 stocks. No one could turn it off. The firm was dead within days, acquired by Getco for salvage. The lesson was simple: automated trading systems need an off switch. That lesson was available for fourteen years before Breeden's speech.
After Knight Capital came the 2010 precedent. On May 6, 2010, the Dow Jones Industrial Average fell 998 points in 36 minutes when Waddell and Reed's algorithm dumped $4.1 billion in E-Mini S&P contracts. The algorithm was programmed to sell at 9 percent of the prior minute's volume regardless of price. Regulators responded by implementing single-stock circuit breakers for all S&P 500 constituents, halting trading for five minutes on any stock that moved more than 10 percent. That mechanism, refined over 16 years, is the architecture Breeden is proposing to reinvent for AI.
The Bank of England is not acting alone. Project Logos, a collaboration between the BIS Innovation Hub, the Bank of England, and Germany's Bundesbank, is running simulations of LLM-based agents operating as portfolio managers in a synthetic financial market. It is the first known attempt to stress-test an agentic financial system at scale before it operates in live markets. The project is designed to identify which elements of agentic AI architecture drive herding.
But the herding is already legible without simulation. Samsung and SK Hynix together constitute approximately 58 percent of the KOSPI index by market capitalization. When one sector moves, the index moves. When the index moves enough, the circuit breaker fires. The concentration is the herding. It does not require artificial intelligence. It requires only that two companies in the same supply chain represent more than half the national equity market.
Breeden also proposed enhanced recovery arrangements, a mechanism by which one bank could take over another's core functions during a disruption. She cited Ukraine's Power Banking program, launched in 2022 during the Russian invasion, in which banks maintained a shared branch network with backup power and internet connectivity to keep the financial system running during infrastructure blackouts. The analogy is revealing. Power Banking was designed for a shooting war. The Bank of England is borrowing wartime infrastructure for a risk that has not yet produced a single documented incident of AI-driven market failure.
The Cambridge survey found that software engineering is the financial industry's most mature AI application, with 42 percent of firms at full deployment. Most agentic AI deployments concentrate in lower-risk operational tasks: research synthesis, data processing, document review. The AI agents that finance firms actually use do not make trading decisions. They summarize earnings calls and draft compliance reports. The kill switch is being designed for a population that does not yet exist in meaningful numbers.
The deeper pattern is regulatory displacement. The risk the Bank of England can see and name is AI herding. The risk it cannot name is market concentration so severe that a single American corporate announcement triggers a sovereign circuit breaker eight thousand miles away. The KOSPI has no AI herding problem. It has a Samsung and SK Hynix problem. No kill switch fixes that.
Breeden said something else in Sintra that deserved more attention than the kill switch. She said AI had produced a step change in cyber capability, and that cyber resilience was one of the Bank's closest financial stability concerns. That risk is real, specific, and already producing incidents. It does not need a new regulatory framework. It needs enforcement of the existing one. But enforcement is invisible. Kill switches make headlines.
In 2010, the kill switch was a response to a crisis. In 2012, its absence destroyed a firm. In 2026, it is a proposal in search of a problem. The circuit breakers already exist. They fire regularly. Every time, the trigger is human. The Bank of England is building a second kill switch for the machines, next to the one that already works for the humans. When both populations herd in the same direction at the same time, neither switch will know whose hand to stay.