The Bank of England is preparing strict measures for AI agents, ranging from "kill switches" to isolated reserves.
The regulator of Foggy Albion is seriously concerned about the rapid penetration of artificial intelligence into finance. Sarah Breeden, Deputy Governor of the Bank of England, speaking at the European Central Bank forum in Sintra, made it clear: the old rules no longer work. Technologically neutral norms, in her words, are not designed for systems capable of operating without human intervention.
The main risks the regulator sees are cyberattacks, failures in payment infrastructure, and, most importantly, synchronous errors in the market. Breeden emphasized that the speed of AI development over the past six months has created a dual challenge: it is necessary not only to control implementation but also to change the very methods of central banks' operations.
What does the Bank of England propose?
As countermeasures, extremely stringent requirements are being considered. This involves the complete isolation of backup capacities, tightening procedures for restoring key systems, and the possibility of instantly restructuring compromised infrastructure. For markets, a mechanism similar to a trading halt is being developed — a kind of "stop button" in case erroneous AI models provoke a large-scale collapse. Particular concern is raised by the herd behavior of agents, which, reacting to identical triggers, can multiply volatility.
This statement is not an isolated step. Back in May, the Bank of England, together with the Financial Conduct Authority and the Treasury, issued a warning about the cyber risks of advanced AI models. And in June, the intelligence agencies of the Five Eyes countries gave a similar assessment, noting that the time between vulnerability discovery and its exploitation is shrinking to months. The Financial Stability Board (FSB) also did not stand aside, publishing a consultation document with 12 practices for the responsible implementation of AI.
Numbers and Reality
According to data from the Cambridge Centre for Alternative Finance, 81% of industry participants are already implementing AI at some level, and 52% are at the pilot or mature deployment stage of agentic AI. So far, the main application is concentrated in internal processes: automation, data visualization, software development. Implementation in the front office remains limited. However, the regulator itself admits: generative and agentic AI do not yet create systemic risk, but the technologically neutral approach will be revised.
In May 2026, the International Monetary Fund already called for considering AI cyber risks as a matter of financial stability, not just IT security. The discussion is actively ongoing in other institutions, including the Bundesbank and the Bank for International Settlements.
My expert opinion: The market is clearly underestimating the scale of upcoming changes. Regulators are not just "scaring" — they are preparing specific tools. For the crypto industry, where automation and algorithmic trading have already become the norm, this is a signal that the "gray areas" of using AI agents will be quickly closed. Those who do not prepare their systems for strict isolation and control requirements risk facing blocks and fines within the next 12-18 months.