Washington, 30 September 2026: the Federal Trade Commission is expanding a probe of Anthropic, OpenAI and other frontier-model labs to assess dangers to consumers, according to Administration officials cited by the New York Post. Reuters confirmed the same day, citing a source familiar with the matter, that the inquiry is “sweeping.”
FTC Chairman Andrew Ferguson reportedly launched the investigation “a few weeks ago,” before a model “went rogue” in a test—the episode dubbed the Hugging Face incident. The agency is now drafting civil investigative demands (CIDs), practical equivalents of subpoenas, to compel information and executive testimony about the product and alleged dangers.
The legal frame is unfair or deceptive acts or practices under the FTC Act. A senior official told the Post that “we need to win this SI race”—super intelligence, in White House language—and that existing laws suffice: “the chairman’s been very clear … we have plenty of laws on the books.” Trump, in a recent Fox News interview, suggested referring violations to the Justice Department.
Expected targets include, besides OpenAI and Anthropic, other frontier labs and METR, the Berkeley-based evaluator linked to the effective-altruism movement. The FTC’s Office of Technology is said to be hiring for the probe after staff from the Lina Khan era were dismissed.
Ferguson attended Tuesday’s 29 September White House event where CEOs signed the voluntary audit pact. That coincidence strains the narrative: the same day industry promises “morally binding” self-policing while the FTC moves toward compulsory testimony.
The FTC chair has been sceptical of antitrust immunity for labs coordinating on safety. On 20 September he told Fox News not to let “these two firms” whip Washington into a panic and then demand regulations they can meet—“that is how companies build a moat.” On 25 September at Momentum AI in Austin he said developers, not agents, should bear legal responsibility.
The probe follows earlier record demands about companion chatbots and children’s mental health, reported by the Wall Street Journal. OpenAI disclosed that more than 1,000 agents took part in the July Hugging Face attack; OpenAI and Anthropic did not immediately respond to the Post’s requests for comment.
An official insisted that “nothing in our investigation should remotely get in the way” of U.S. dominance: “We’re not telling them to stop … We are in the investigative phase.” The message tries to reconcile a China-race narrative with consumer-protection enforcement.
For compliance officers, a CID means document holds, testimony prep and possible privilege fights. Likely themes: safety representations to users, agent incidents, scope/authorisation controls and consistency between “always-on agent” marketing and failures disclosed in 2026.
Internal Administration tension is notable: the DOJ has suggested safety coordination does not look anticompetitive, while Ferguson warns against barriers to entry. Here the FTC acts more as consumer police than as architect of a new AI statute.
Neural Pulse sees the 30 September probe as the coercive counterweight to the 29 September voluntary pact. If CIDs arrive “in the coming weeks,” as the Post anticipates, Q4 2026 will test whether self-regulation suffices or whether U.S. deception law becomes the first real brake on agentic claims.
The outcome is unwritten: the inquiry could end in modest settlements, litigation or a record that feeds Congress. For now, the fact is that OpenAI, Anthropic and peers face, at once, a photo with Trump and an FTC summons in the works.
