AI Trading’s Real Risk Is the Feedback Loop, Not One Bad Model

The most dangerous AI trader may not be the one that makes a bad decision. It may be the one that makes a reasonable decision after watching thousands of other machines do the same thing. That is the regulatory problem hiding behind the current enthusiasm for agentic finance: markets are not merely getting new tools. They may be acquiring new participants whose reactions can amplify one another at machine speed.

The Commodity Futures Trading Commission is putting that possibility on the agenda. The agency’s Innovation Task Force has announced a Frontier Forum session for October 28, 2026, focused on AI in derivatives markets and the rise of agentic finance. The session is part of a wider regulatory conversation, not proof that autonomous trading systems have already taken over the market. But the choice of subject matters. It suggests that regulators are beginning to treat AI as a question of market structure, not simply software procurement.

The CFTC is right to move from models to systems

For years, the natural regulatory question was: Is this algorithm accurate, fair and compliant? That question still matters, but it is too narrow for systems that can interpret information, choose actions and potentially adjust their behavior with limited human involvement.

The CFTC’s own earlier work points in this direction. In 2024, its Technology Advisory Committee recommended attention to human involvement, governance, data quality, privacy and risk-management frameworks designed for AI-specific risks. The committee also warned that generative AI could create instability when firms do not fully understand or control the data and strategies involved, as the CFTC’s report explains.

That is a useful foundation, but the next step is to stop treating each deployed model as an isolated object. A trading agent operates inside a chain: data enters, a model interprets it, an action reaches a market, other systems react, and the resulting data becomes part of the environment for future decisions. The danger is located in that loop.

A model can be individually defensible while the combined behavior of many models becomes destabilizing. If several agents identify the same signal, pursue the same position or retreat from risk simultaneously, their actions can turn a small movement into a larger one. This is not a claim that such a scenario is inevitable. It is an argument that testing one model at a time cannot reveal all the risks of a connected market.

“Human oversight” must mean an actual control, not a slogan

Regulatory language often treats human involvement as a reassuring final checkpoint. But a person who receives an automated alert after a position has already changed rapidly is not necessarily exercising meaningful control. They may simply be observing an event they cannot reverse quickly enough.

The practical standard should be intervention capacity. Firms should be able to explain what an agent is authorized to do, which decisions require approval, how its activity can be paused, and who is responsible when the system behaves outside its expected boundaries. These are opinions about what effective governance should require, not details announced by the CFTC in the documentation. Yet they follow directly from the agency’s concern about firms lacking full understanding or control over AI-enabled trading and risk programs.

The same principle applies to vendors. If a financial institution relies on a third-party model whose training data, weighting or updates are opaque, responsibility cannot disappear into the supply chain. Intellectual-property protection may limit what a provider reveals, but it should not prevent a regulated firm from managing and reporting its own trading and clearing data. If it does, the arrangement is not merely a technology issue; it is a governance gap.

Testing should include the market’s reaction to the agent

Back-testing a strategy against historical data is not enough when the strategy changes the conditions it is trying to exploit. Regulators and firms need testing that asks how systems interact: whether agents converge on similar behavior, how quickly they abandon positions, and what happens when data is incomplete, manipulated or simply wrong.

The CFTC’s 2024 discussion identified trading, risk management, compliance, customer service and back-office operations as relevant AI use cases. That breadth is important because instability may not begin with an obvious trading instruction. A flawed risk assessment, a bad compliance classification or an automated customer interaction could alter decisions elsewhere in the chain.

This is where the CFTC’s forum could become more than a listening exercise. The agency should push participants to describe measurable failure conditions and practical shutdown procedures, rather than accepting broad promises about responsible innovation. The public needs to know not only whether an agent performs well under normal conditions, but also how quickly a firm can detect a problem and contain it.

The winners will be firms that can prove control

Rules for agentic finance should not aim to ban automation. AI already has established uses in financial markets, including surveillance, fraud detection, back-testing and risk management, according to the CFTC’s earlier report. Properly designed systems can reduce manual work and identify unusual activity faster than people can.

The better regulatory target is unaccountable autonomy. Firms that maintain clear records of model changes, permissions, data sources, decisions and interventions should have an advantage over those that treat an AI agent as a black box with a trading account. That approach would also benefit serious developers: predictable expectations are easier to build around than vague demands to “use AI responsibly.”

The CFTC’s October 28 forum is therefore a signal, not a solution. Its importance will depend on whether the conversation moves beyond the capabilities of individual models and addresses the collective behavior of machines connected to the same market. In our view, regulators should govern the feedback loop first: require accountability for every decision path, test interactions under stress, and make intervention real before autonomous finance becomes ordinary infrastructure. By the time a machine-speed feedback loop is visibly damaging markets, oversight will already be late.