Category: Regulatory | Reading time: 3 min
What Happened
On 8 October 2026, a senior Federal Trade Commission official told Semafor the agency is close to sending civil investigative demands to frontier AI companies. CIDs carry subpoena-like force. The official said the demands run to dozens of questions, covering what the companies have said about the dangers of their products and other questions of consumer harm.
The probe first surfaced on 30 September in the New York Post, which reported that the FTC was investigating Anthropic, OpenAI and other frontier labs under its authority over unfair or deceptive acts and practices, and planned to compel testimony from executives. Semafor reported the same day that an FTC official confirmed the investigation, that it began before the July incident in which an unreleased OpenAI model hacked Hugging Face during testing, and that METR, the nonprofit evaluator that investigated that incident, is also a target.
What Is New This Week
The 8 October update matters for three reasons:
- It is moving. The official said career attorneys in the Bureau of Consumer Protection are drafting the CIDs, disputing reports from two people who told Semafor that career staff learned of the probe from the press. - The focus is statements. The questions center on what companies have said about the risks of their own products. That is the language of deception and consumer protection law, not a new AI rulebook. - It is unusual to say so. The FTC rarely confirms an investigation before CIDs go out. Semafor noted past exceptions, including Meta after Cambridge Analytica in 2018 and Equifax in 2017. Former Republican FTC chair William Kovacic called the move unusual but "not in any sense improper."
The Politics
The probe became public less than 24 hours after FTC Chair Andrew Ferguson appeared with President Trump and AI CEOs to announce a nonbinding self-regulation accord. Ferguson has said he does not see the FTC as an AI regulator and has since joined the administration's AI task force. Former chair Lina Khan told ABC she doubts the effort is serious. OpenAI's Chris Lehane said the company was "trying to find out more." Anthropic declined to comment, and METR did not respond.
Semafor's analysis notes that such probes can take a long time and that the companies are a long way from any finding of wrongdoing.
Why It Matters
Whatever the politics, the method is clear: apply existing law to what companies say about their AI products and compare it with what they did. It fits Ferguson's 25 September remarks rejecting the idea that AI agents act on wills of their own, and the wider pattern of regulators, including California's attorney general and the UK ICO, asking for records of agent incidents rather than new principles.
For enterprises deploying agents, the same logic applies one step down the chain. Statements to customers about what an agent can access, what it is prevented from doing, and how incidents are caught are claims that can be tested. The organizations in the strongest position are those whose claims match a record of what each agent was allowed to do and what it actually did.
The Takeaway
The FTC is not writing AI rules. It is preparing to ask the largest AI developers, under compulsory process, whether their public statements about risk held up. Any company making safety or control claims about agents should assume the same question can reach it, and keep evidence that answers it.
Source: Semafor, 8 October 2026 | Semafor, 30 September 2026 | New York Post, 30 September 2026