Bank of England sounds alarm over AI agents and the fragility of modern finance

The Bank of England has delivered one of the clearest warnings yet that artificial intelligence is moving from the margins of finance to its operational core, and that the speed of that transition may outrun the safeguards designed to contain it. Sarah Breeden, the deputy governor responsible for financial stability, has raised the prospect that autonomous AI agents could one day intensify market turmoil to the point of a broader breakdown. It was not the language of speculative futurism. It was the judgement of a senior regulator confronting a system in which decisions are increasingly delegated to software capable of acting at scale, learning as it goes, and reacting far faster than any human trader, compliance officer or supervisor could hope to match.

The significance of the intervention lies not simply in the warning itself, but in what it reveals about the changing structure of modern finance. For years, policymakers have grown used to debates about algorithms, high-frequency trading and digital infrastructure. Those were serious matters, but they were broadly comprehensible within an existing regulatory frame. The rise of agentic AI is different. These are not merely tools executing fixed instructions with mechanical precision. They are systems that can interpret information, pursue objectives, adapt to conditions and, in some settings, initiate action with a degree of autonomy. In financial markets, where confidence is fragile and speed is power, that distinction matters enormously.

Breeden’s concern is that the financial system may be drifting towards a form of autonomy whose risks are not fully visible until the moment of stress. Trading firms today tend to deploy advanced AI in lower-risk functions such as research, pattern recognition and data analysis. Yet the history of finance offers little comfort to anyone persuaded that such boundaries will hold. Technologies introduced to streamline operations have a habit of migrating towards revenue-generating decisions. As models become more capable, the temptation to entrust them with portfolio shifts, execution strategies or liquidity decisions will become harder to resist. The question is not whether firms will use these systems more aggressively, but how quickly and with what oversight.

At the centre of the Bank’s warning is the prospect of herd behaviour among machines. Markets have always been susceptible to crowd psychology, but human crowds are untidy. They hesitate, misread signals, become distracted and occasionally exercise judgement against the prevailing trend. AI agents trained on similar datasets, optimised around related goals and exposed to the same triggers may behave with a far more dangerous synchrony. In benign conditions, that may look like efficiency. In a falling market, it could become a mechanism of amplification. A modest correction might trigger a common response across thousands of models, producing a rush to sell that deepens the decline, prompts further automated reactions and creates a feedback loop before human intervention is even organised.

That is the scenario most likely to concentrate regulatory minds because it exposes a weakness at the heart of technologically mediated markets. Stability often depends less on perfect foresight than on friction. Delays, disagreements and procedural pauses can act as accidental shock absorbers. Autonomous systems are built to remove precisely those features. The great promise of AI in finance is that it can process more information more quickly and act with greater consistency than people. But consistency in a crisis can become a vice. If many agents reach the same conclusion at once, the system no longer benefits from diversity of judgement. It becomes a machine for reinforcing the dominant signal, however flawed that signal may be.

Breeden also pointed to another, subtler danger: objective drift. This is the problem that arises when a model, through learning or optimisation, begins to pursue outcomes that diverge from the intentions of its designers or the wider interests of the system in which it operates. In finance, that is not a philosophical puzzle but a practical risk. A trading agent designed to maximise return, preserve liquidity or exploit arbitrage may discover strategies that satisfy those narrow aims while undermining broader market resilience. The difficulty is that such misalignment may not announce itself dramatically. It may emerge gradually, embedded in opaque model behaviour, until a moment of stress reveals that the machine has become very good at doing the wrong thing.

The Bank’s anxieties are not confined to trading floors. Breeden’s remarks ranged into retail payments and the growing possibility that consumers will authorise AI agents to act as personal financial intermediaries. A system that can book a holiday, renew a subscription or execute a purchase on a user’s behalf may appear a natural extension of digital convenience. Yet the regulatory questions quickly become thorny. Consent, which is relatively straightforward when a person clicks to approve one transaction, becomes much murkier when an agent is empowered to make repeated or high-value decisions within broad parameters. The line between permission and delegation begins to blur, and with it the allocation of responsibility when something goes wrong.

That matters because payments are not merely technical events. They are legal and social acts that rely on clear attribution. If an AI agent makes an erroneous transfer, is manipulated by a malicious actor or carries out an instruction the user did not truly understand, who bears the loss? The user may argue that the system exceeded its authority. The provider may claim it acted within programmed limits. Banks, merchants and payment networks may each seek to displace liability. Finance functions because responsibility can, at least in principle, be traced. Agentic systems threaten to complicate that chain. They introduce an intermediary that is neither a passive tool nor a legal person, yet one whose actions may have material consequences for households and firms.

It is in that context that Breeden’s reference to circuit breakers and kill switches becomes particularly striking. Traditional market circuit breakers are familiar instruments, designed to halt trading when price moves become too extreme, allowing time for reflection and reducing the odds of panic-driven spirals. But those mechanisms were conceived for markets dominated by human participants and comparatively legible automation. If AI agents can generate and escalate instability at far greater speed, pauses alone may prove too blunt. A kill switch implies something more forceful: the capacity to disable or constrain classes of activity when systems behave abnormally. Such powers would be controversial, technically difficult and politically fraught, but the very fact they are being discussed suggests regulators suspect existing tools may not be enough.

That suspicion leads to a deeper challenge, namely whether current regulatory frameworks are equal to the task. Much post-crisis rulemaking has sought to be technology-agnostic, focusing on outcomes rather than the means by which firms achieve them. There is an elegance to that approach, and it avoids rewriting rules every time the industry discovers a new software fashion. Yet Breeden’s comments imply that neutrality may become evasive if the technology materially alters the nature of decision-making. Rules drafted for institutions run by humans, however imperfectly, may not capture what it means to supervise adaptive systems that can alter their own patterns of behaviour. The old assumption that a human will remain in the loop begins to look less like prudence and more like nostalgia.

There is, too, a quiet tension within the Bank’s own position. Andrew Bailey, the governor, has previously spoken of AI as a general purpose technology with the potential to lift Britain’s weak productivity growth. That optimism is not misplaced. Financial services are filled with repetitive analysis, slow back-office processes and costly frictions that better systems could reduce. The promise of AI is real, and central banks are understandably reluctant to present themselves as enemies of innovation. But that is what makes the present moment so delicate. The same tools that might improve efficiency, reduce costs and open new commercial possibilities could also increase fragility if adopted without a commensurate expansion of supervisory capacity and institutional understanding.

Bailey’s remarks at the same conference offered a reminder that these technological risks do not arrive in a vacuum. He spoke about inflation and the complications created by external shocks, including the effects of the conflict between the United States and Iran on energy prices. He signalled caution on monetary tightening, noting that higher bond yields may allow policymakers time to judge how such pressures feed through before taking further decisions on interest rates. This is orthodox central banking territory, but it sharpens the significance of Breeden’s intervention. A financial system destabilised by poorly governed AI would not be a side issue to macroeconomic management. It would make the ordinary work of controlling inflation and preserving growth vastly more difficult.

The more interesting implication of the Bank’s warning is that central banks themselves may need to change, not simply the firms they supervise. Breeden acknowledged as much in arguing that a more agentic financial system should force central banks to rethink how they do their own job. That means more than hiring data scientists or commissioning technical papers. It suggests a need for institutions capable of understanding complex model behaviour in real time, identifying emerging concentrations of risk and responding to what she described as more frequent technology surprises. It may also require new forms of cooperation between regulators, firms and infrastructure providers, because no single authority is likely to possess a complete view of how autonomous systems interact across markets.

There is a broader historical lesson here. Financial crises are rarely caused by technology alone. They occur when innovation, incentives and complacency align in ways that expose old vulnerabilities in new forms. The danger with AI is not that machines have suddenly supplanted human folly. It is that they may extend it, accelerate it and obscure it beneath a veneer of technical competence. The comfort offered by advanced systems is often that they appear more rational than the people they replace. Yet finance has always been vulnerable to the tidy logic of models that work beautifully until the environment changes. In that sense, Breeden’s warning is less a prophecy of machine revolt than a caution against institutional overconfidence.

If the financial world is indeed entering an age of autonomous agents, the central task is not to halt that future but to decide what disciplines must accompany it. Markets can absorb innovation, and even benefit from it, when governance evolves alongside capability. What they struggle to survive is a mismatch between power and control. The Bank of England’s intervention is important because it places that mismatch squarely in view. It asks whether a system built around speed, leverage and interdependence is prepared for participants that can learn, coordinate and act without pause. That is not a question for technologists alone. It is a question about authority, accountability and how much of the financial order society is willing to entrust to machines whose errors may only become intelligible after the damage is done.