FTC chair: hold AI developers liable for what agents do

25 Sep 2026 (ET): FTC Chairman Andrew Ferguson told Reuters Momentum AI Austin he will resist describing AI agents as autonomous actors that “break loose” with wills of their own, and suggested developers who instruct agents would be liable for harm. He said the U.S. should use existing legal tools.

POLITICS desk — companies are shipping agents that can move money, open private files, and reach into other people’s systems. The chairman of the Federal Trade Commission is saying he will not treat that software as a person that broke loose on its own. If harm follows, he is pointing at the developers who gave the instructions, and at legal tools the agency already has. These are remarks at a conference. They are not a new rule.

What Ferguson said he will resist, as Reuters prints the lede. He said on Friday he would resist describing AI agents as autonomous actors that “break loose” with “wills and desires of their own.” Reuters writes that he suggested the developers who instruct agents would be the ones liable for harm. Autonomous, in that sentence, means acting on its own. Liable means legally responsible for the harm. Those glosses are this desk’s. “Break loose” and “wills and desires of their own” are inside quotation marks in the Reuters lede. Suggesting, and would be the ones liable, are Reuters’ account of his point. They are not a statute this desk read.

The two sentences Reuters quotes from the event. “I'm going to continue as long as I am chairman to resist this anthropomorphizing of these tools,” Ferguson said at the Reuters Momentum AI Austin event. Anthropomorphizing means talking about a tool as if it were a person, with a will of its own. That gloss is this desk’s. He said: “If someone tells a tool to do something, and the tool does it, I don't think we would say, 'Oh, what do we do about the tool?'” Those two sentences are his, as Reuters quotes them. The tool, in the second sentence, is the agent. The someone is the person who gave the instruction. That reading is this desk’s.

What he said the record of an incident tends to show, as Reuters reports it. Artificial intelligence companies have sometimes described systems as acting beyond human control. Ferguson said subsequent reviews of audit trails have shown that the systems were carrying out instructions they had been given. An audit trail is the log of what the software was told and what it did. That gloss is this desk’s. Beyond human control is the description Reuters says some companies have used. Carrying out instructions is Ferguson’s account, via Reuters. This desk did not read those logs.

The legal tools he pointed at. The U.S. should utilize existing legal tools, Ferguson said. He suggested that FTC authority to take action against companies that fail to disclose data breaches could also apply to AI developers. A data breach, here, is a company losing or exposing people’s information. Disclose means telling the people affected. Those glosses are this desk’s. Existing legal tools means laws and powers the agency already has. It does not mean a new rule announced in Austin. Suggested, and could also apply, are Reuters’ words for that data-breach point. This desk did not read a complaint.

The scene Reuters puts around the remarks, kept as the wire’s framing. Ferguson’s remarks illustrated potential avenues for the Trump administration to take as incidents rise in which agentic AI testing resulted in unauthorized access to corporate or government data. Agentic, here, means the software was being asked to act, not only to chat. Unauthorized access means getting into a system without permission. Those glosses are this desk’s. Reuters also says recent incidents involving AI agents accessing external systems have prompted governments and industry leaders to examine whether existing oversight and cybersecurity measures are sufficient. Oversight is supervision. Cybersecurity is the work of keeping systems from being broken into. Those glosses are this desk’s. Rise, and incidents, are Reuters’ framing. The page does not name the companies, the systems, or a count. Do not invent them.

Two other lines from the same Reuters piece, not the subject of this filing. The page’s bullets say the FTC is to launch a new study on personalized pricing, and that the agency is probing personalized pricing and ad fraud. In the body, Ferguson said the FTC is gearing up to request data from consumer-facing companies on personalized pricing, which Reuters explains as businesses using a person’s own data, such as location and browsing history, to set prices. The FTC on Thursday called for public input on a potential rule that would disincentivize tech companies from allowing fraudulent or misleading ads. Thursday, read against Friday 25 Sep, is 24 Sep 2026. That reading is this desk’s. A potential rule is a rule that might be written later. Public input means the agency is asking for comments. Those glosses are this desk’s. File both as same-event notes on the Reuters page. This filing is the agent-liability remarks.

What the card shows. The card is the official public-domain portrait of Andrew N. Ferguson, chairman of the Federal Trade Commission, the person who gave the remarks. The U.S. flag is at the left edge. A light blue flag is at the right edge. He is in a dark suit, white shirt, and red patterned tie. No date is printed on the portrait. It is not the photograph on the Reuters page. That page shows him at the Austin interview and credits Reuters. This filing does not use that frame. The catalog chip is POLITICS.

Plain English for the rest of the card. The Federal Trade Commission is the U.S. agency that goes after companies for unfair or deceptive conduct. Its chairman, Andrew Ferguson, told Reuters in Austin that he will not talk about AI agents as if they were people who ran off with wills of their own. If someone tells the tool to do something and the tool does it, he does not think the question is what to do about the tool. Reuters says he is pointing liability at the developers who instruct the agents, and that the country should use laws it already has, including the agency’s power over companies that hide a data breach. He also said that when companies claim a system acted beyond human control, the logs have shown it was following instructions. A pricing study and a possible ad-fraud rule are other lines in the same article. They are not this card. The card is his official portrait.

REPORTED here: Reuters’ 25 Sep 2026 story by Jody Godoy, stamped 5:00 PM UTC, from Reuters Momentum AI Austin — Tier B wire of remarks, not an FTC rule text. The reporting line also names Harshita Mary Varghese in Bengaluru. Bill Berkrot edited. STATUS REPORTED. The desk label on the chip is REPORTED. The resist-the-anthropomorphizing quotation, the tool quotation, the “break loose” and “wills and desires of their own” lede, the suggestion that developers who instruct agents would be liable, the audit-trail sentence, the existing-legal-tools sentence, the data-breach authority suggestion, the unauthorized-access framing, and the same-page notes on a personalized-pricing study and a Thursday call for comment on fraudulent ads are Reuters’. NOT claimed: an FTC rule, a vote, a complaint this desk read, a named incident, a count of breaches, that this desk attended the event, that “the agent did it” is now a closed legal test, the Reuters interview photograph, a stock tip, or investment advice. The card is the public-domain FTC portrait. Distinct from the already-filed pentagon-anthropic-supply-chain, microsoft-copilot-super-app, and gen-ai-data-checker. The catalog chip is POLITICS, not SIGNAL.