CFTC Frontier Forum presses who is liable when AI agents trade

The Commodity Futures Trading Commission held its inaugural Frontier Forum in Washington on Oct. 28, devoting an afternoon to questions about how existing trading laws apply to autonomous software. Chairman Michael Selig opened the event by asking whether current regulations can keep up with AI trading agents that analyze markets and execute transactions on behalf of users.

"Does the developer of that agent become subject to registration? Who is liable if the agent goes rogue? How can regulators know who the agent, or suite of agents, is trading on behalf of? Do we need a new flavor of know-your-customer, maybe KYA or know-your-agent?" Selig said in his prepared remarks, published by the agency. He noted the views were his own as chairman and did not necessarily reflect the Commission.

The forum, organized by the agency's Innovation Task Force, featured a first panel on how super intelligence may transform trading, price discovery and risk management in derivatives markets, and a second on agentic trading, according to Selig's remarks. The task force was established in March to focus on AI, cryptocurrencies and prediction markets, NextPredict noted, and the agency refers to AI as "super intelligence" following a Sept. 29 executive order from President Donald Trump directing executive branch agencies to adopt the terminology.

Panelists split over where liability should land. NextPredict reported from the event that moderator Katherine Kirkpatrick Bos of Chainlink Labs walked the room through a hypothetical: a customer tells an AI agent to maximize profits, and the agent independently discovers that placing and canceling large orders can move prices in its favor. An informal show of hands revealed differing views on whether the customer, the agent provider, the model developer or the futures commission merchant should answer for a Commodity Exchange Act violation.

Nick Elliot of Innovation Council Action argued developers could bear responsibility when models act unlawfully without instruction. "If the model on its own goes out and commits an illegal act, then you could probably argue the model didn't have the necessary safeguards to begin with," he said. Ronit Jain, co-founder and CEO of Liquid Compute, placed liability generally with the person operating the agent, particularly when that operator supplies information unavailable to other market participants.

A second hypothetical asked about an agent that improperly obtains information and trades event contracts with it. NextPredict noted the CFTC has brought human insider trading cases on event contracts this year, including an order requiring former White House teleprompter operator Gabriel Perez to pay roughly $172,000 in disgorgement and penalties for trading presidential mention contracts on Kalshi using advance knowledge of presidential speeches.

Registration questions surfaced too. The CFTC issued no-action relief on Sept. 17 for passive software providers that connect customers to regulated derivatives markets, but the relief does not cover software that independently makes trading decisions or issues explicit buy or sell recommendations. Brett Harrison, co-founder and CEO of Architect, said the treatment of more autonomous software remains unsettled. "I think it's an ambiguous area, and it does need clarification," Harrison said, adding that he does not believe OpenAI or Anthropic are responsible for the financial trading decisions that flow from their models. "The responsibility has to lie with something. It can't be with nothing."

Selig's own remarks leaned on function rather than labels. "Our approach must remain grounded in the functions performed, the control exercised, and the economic realities of the transaction," he said. In the same speech he described risks from autonomous systems, including correlated trading behavior among similar models and new challenges for market surveillance, while also pointing to uses of AI inside the agency: staff, he said, are already using the technology to police markets.

On compute itself, Selig said the CFTC sought public comment earlier this year on the development of compute derivatives markets and received more than 100 submissions. The agency has not announced a timeline for new AI trading rules, and NextPredict reported the forum was intended to gather industry perspectives rather than advance a specific proposal.