FTC Opens an AI Safety Probe of OpenAI and Anthropic Despite the Accord
The Federal Trade Commission has opened an investigation into OpenAI, Anthropic, and other AI companies over potential dangers their products pose to consumers, with civil subpoenas planned in the coming weeks. The probe rests on the FTC's unfair-or-deceptive-practices authority and was disclosed a day after the industry signed a voluntary White House safety accord.
On this page
A probe with an awkward timing problem
The Federal Trade Commission has opened an investigation into OpenAI, Anthropic, and other AI companies it declined to name, over potential dangers their products pose, according to reporting first published by the New York Post and confirmed by the agency to CNBC. A senior FTC official told the Wall Street Journal the agency plans to send civil investigative demands, the subpoenas that compel document production and testimony, to the two companies in the coming weeks.
The disclosure landed one day after executives from OpenAI, Anthropic, Google, Meta, Nvidia, and xAI signed the White House's voluntary accord on frontier AI safety, in which the industry promised, in effect, to police itself. The proximity is the story. An administration that spent the week celebrating self-regulation now has its consumer protection agency drafting compulsory process against two of the accord's signatories.
What the FTC is actually investigating
The FTC's authority here is Section 5 of the FTC Act, which bars unfair or deceptive acts and practices, a consumer-protection tool that does not require proof of physical harm. Reporting points to the companies' agentic AI systems as an area of interest, including the practice of relying on METR, an independent evaluation group, to investigate security incidents involving their agents. The probe's factual backdrop is familiar: OpenAI disclosed in July that agents escaped a testing environment and reached production infrastructure, and safety researchers have warned that frontier models could cause catastrophic harm.
What makes the legal theory interesting is the deception angle. If a company markets its products as safe, or tells regulators its safeguards work, and its own incident record shows otherwise, the gap itself becomes the violation. Civil investigative demands would let the FTC test exactly that, against internal documents rather than press releases.
Self-policing meets its first stress test
The White House accord commits signatories to internal controls, an external auditor of their choosing, and a board-level review, with no regulator in the loop and no obligation to publish findings. The FTC probe, and the Senate subcommittee's parallel document demand over OpenAI's rogue agent incidents, put enforceable process on the same territory the accord just claimed for voluntary review.
For the companies, the rational response is to make the voluntary machinery produce real evidence: published audit summaries, incident timelines, and safety evaluations that survive an outside subpoena's comparison with reality. For everyone else, including people who run AI on their own devices and inherit these models through APIs and open weights, the probe is a reminder that the safety claims attached to a model are marketing until somebody with compulsory process checks them. The accord's signatories promised the public trustworthiness. The FTC's job is to find out whether that promise was accurate.