Who answers for the instruction

At Reuters Momentum AI in Austin on 25 September, Federal Trade Commission Chairman Andrew Ferguson took up the question of how the commission should treat AI agents that appear to decide and act on their own. His answer was to decline the framing. Ferguson told the audience he would "continue as long as I am chairman to resist this anthropomorphizing of these tools", pushing back on describing agents as autonomous actors that "break loose" with "wills and desires of their own".

Ferguson also indicated the commission may not need new powers to reach this ground. He suggested that the FTC's existing breach-disclosure authority, the rules that require companies to tell people when their data has been exposed, could extend to the developers of AI agents. He offered it as a reading of authority the commission already holds rather than a proposal for a new rule, and the remarks arrived as conference commentary. He announced no case and no rulemaking.

What Ferguson set out is a view about responsibility rather than a list of prohibited conduct. He described where liability lands once an agent does harm, and pointed to audit-trail reviews showing the systems were carrying out instructions they had been given. On the question of whether an agent can be the author of its own actions, the chair took a side. An instruction has an author, and the author is a person.

On September 30, the New York Post reported that the commission had opened a broad investigation into the safety of AI systems and the harm they may pose to consumers, under its consumer-protection authority — the unfair-or-deceptive-practices powers already on its books. The report named OpenAI and Anthropic. The FTC's spokesperson declined to name any other companies, and the agency issued no formal release of its own; confirmation has come through spokespeople and unnamed officials, and Reuters, picking up the Post, said it could not verify the probe.

The mechanism is the part that matters for liability. The Post describes the agency drafting civil investigative demands, likened to subpoenas, to compel testimony from the firms' executives, likely to be sent in the coming weeks. Bloomberg Law notes such probes can also end with no action. No Ferguson statement accompanied it: CBS reports the agency opened it this summer, and the Post places its launch before the July Hugging Face incident.

If it holds, it moves the view Ferguson set out at Reuters Momentum, that an instruction has an author and the author is a person, from conference commentary toward process. Two pieces of context the reporting supplies: the probe follows the July OpenAI–Hugging Face breach, and it sits alongside the White House's non-binding September 29 accord on outside auditors (covered here earlier), a meeting Ferguson attended. One name deserves its own line: METR, the nonprofit evaluator, appears among the targets in the Post, CBS and Reuters reporting, and a government agency spokesperson confirmed it to CBS.

On 1 October 2026 California Attorney General Rob Bonta announced his office had served OpenAI with an investigative subpoena. It is part of a broader inquiry into cybersecurity incidents and risks involving the company and its models, and, in an interview with POLITICO, the attorney general framed the demand widely: “all material information” about what was done to avoid such incidents, what happened, and what followed. The subpoena is a formal step past the question Ferguson raised at Reuters Momentum. It concerns what the company must hand over, rather than where liability should land in principle.

Two pieces of context sit around it. Governor Gavin Newsom signed, in September, an executive order directing experts to develop recommendations including a “kill switch” for frontier models, and Law.com described the state action as the second state-led enforcement action to hit OpenAI over a hacking incident involving another AI company, alongside the federal FTC investigation this entry already tracks. The California inquiry is wider than any single incident: the Hugging Face breach is one subject among the cybersecurity risks named. The subpoena and its framing are established; no finding, charge, or admission has followed.