Summary

• Bank of England deputy governor Sarah Breeden has set out how agentic AI could affect cyber risk, markets, and payments.
• The Bank is exploring simulations with the BIS Innovation Hub and Bundesbank, including herding behaviour among AI agents.
• Potential safeguards include circuit breakers or kill switches if faulty AI models threaten market stability.

The Bank of England is examining whether financial markets need new safeguards for agentic AI, including circuit breakers or kill switches that could halt trading if faulty models amplify instability.

Sarah Breeden, deputy governor for financial stability, used a speech at the European Central Bank’s Sintra Forum to argue that AI is reshaping finance at a pace that requires central banks to adapt how they monitor, simulate, and manage risk. Her focus was not only AI adoption inside banks, but the way autonomous systems could affect cyber resilience, markets, and payments.

The Bank is experimenting with the BIS Innovation Hub and the Bundesbank on simulation methods that could test whether aspects of agent design might drive herding behaviour. The work may also examine whether agent objectives can incorporate public policy constraints and whether market-wide guardrails are needed if AI models behave dangerously.

Algorithms are already embedded in finance, from automated execution to quantitative trading and fraud monitoring. Agentic AI changes the supervisory problem because systems can plan, act, interact, and adapt with less direct instruction. During periods of stress, similar models trained on similar data and optimised for similar goals could respond in correlated ways.

Financial stability regulators have always worried about herd behaviour, but autonomous AI could compress the time available for intervention. Human traders, supervisors, risk teams, and market operators already struggle when volatility accelerates. Systems acting at machine speed could make intervention harder, especially if their reasoning is opaque or if their actions spread across several institutions at once.

The kill switch idea therefore sits within a familiar but sharper resilience question. Circuit breakers already exist in financial markets, but AI-driven trading raises new questions about where controls should sit. Safeguards could be embedded in venues, mandated for firms deploying agents, designed into model objectives, or coordinated by market authorities.

Each option carries trade-offs. A market-wide halt can prevent disorderly spirals, but poorly designed controls can also interrupt legitimate liquidity or create incentives for firms to move activity elsewhere. Controls built into individual firms may be more targeted, but they depend on supervisors understanding model design, data sources, and live behaviour well enough to judge the risk.

Payments create a different set of problems. If AI agents begin acting on behalf of individuals or businesses, the financial system will need clearer rules around consent, authentication, liability, and fraud. A system that initiates payments or negotiates purchases may improve convenience, but it also blurs the boundary between user instruction and automated action.

Breeden’s speech places AI governance inside the broader resilience agenda already facing financial firms. Banks and market infrastructure providers are managing cloud dependence, cyber attacks, outsourcing, and critical third party risk. Agentic AI adds a further layer because the underlying models may come from technology providers whose systems influence decisions across many regulated institutions.

Financial firms will need more than model risk paperwork and internal acceptable use policies. Higher risk AI use cases will require live monitoring, scenario testing, rollback mechanisms, audit trails, human accountability, and evidence that systems behave safely under stress. The more autonomous the tool, the less credible it becomes to rely on a human in the loop for every action.

The Bank’s work remains exploratory, but its direction is clear. Regulators are moving beyond the question of whether AI can improve productivity in finance. They are testing whether the financial system’s controls are strong enough for markets where software can act, interact, and fail faster than existing oversight was designed to handle.