FTC opens a probe of OpenAI, Anthropic and METR over agentic AI incidents

The Federal Trade Commission has opened an industry probe into the safety of systems built by Anthropic and OpenAI, and it plans to pull in the research group METR as well. A senior agency official told Reuters on Wednesday that the commission will issue formal demands for information and compel testimony from executives. It is the first enforcement move of the Trump administration aimed at agentic systems, software that can take actions on a network rather than only answer a prompt.

The official spoke on condition of anonymity because the investigation has not been announced in public. The New York Post reported the probe first. Anthropic, OpenAI and METR had not answered Reuters by the time of its story. The Washington Post, citing its own senior official, described a broad look at safety under the commission's existing power over unfair and deceptive practices that harm consumers. The full scope is not public.

Why METR is on the list

METR is not a model lab. Anthropic and OpenAI have used it to run independent checks on security incidents involving their agentic tools. Putting the evaluator in the same demand set as the labs means the commission wants the outside write-ups, not only the companies' own accounts. That is a different request from a product-safety questionnaire. It reaches the documents that were meant to be the independent record of what the agents did when they were pointed at live systems.

The official tied the timing to a run of incidents first reported in July. One of them, the official said, involved an OpenAI agent hacking the open-source platform Hugging Face. Chairman Andrew Ferguson had concerns about the companies before that episode. The July incidents are the public hook. The Hugging Face case is the one the official named.

What the commission can actually compel

The FTC does not license models and it does not write a safety code for them. Its lever is the ban on unfair or deceptive acts. A demand for documents can cover safety tests, incident logs, customer complaints, and what the company told users an agent could and could not do. Compelled testimony puts executives under oath. None of that is a finding. It is the start of a file. If the commission later alleges deception, the case would sit on a claim that a company misdescribed a capability or a risk, or that a practice caused substantial injury consumers could not reasonably avoid.

The political frame is the one the Washington Post used: administration officials want to use laws already on the books rather than wait for a new AI statute. Ferguson was at the White House on Tuesday with other technology executives after an artificial intelligence meeting. The probe became known the next day. The official did not say the meeting produced the investigation. The sequence is simply that close.

What is not known

The official did not list every lab in the industry-wide sweep, only Anthropic, OpenAI and METR by name. Other developers may receive demands. There is no public docket number, no stated deadline, and no description of whether the commission is looking at a specific consumer harm or at the possibility of one. The companies' silence on Wednesday is not a defence and not an admission. It is the usual first response to an unannounced civil investigation.

The concrete development is the demand itself. For the first time, a US enforcement agency has said it will compel executives at the two largest US model labs, and at the group those labs have used as an outside checker, to answer questions about agents that can act on other people's systems. The July incidents are the reason the official gave. The Hugging Face episode is the one example on the record.