The FTC's First Rogue-Agent Probe Names METR Alongside OpenAI and Anthropic. The Auditor Was Supposed to Be the Answer, Not a Subject

The Federal Trade Commission confirmed on Wednesday that it has opened an investigation into OpenAI, Anthropic and other AI developers over the risks their products create for consumers. Reuters, Bloomberg and CNBC each carried the confirmation; the New York Post reported the inquiry first. The agency opened it over the summer and plans to issue formal requests for information and civil investigative demands compelling executives to testify. There is no press release. As of Wednesday the FTC's newsroom listed nothing on the subject, and the agency's position reached the public entirely through reporters.

The third name on the list is the one worth sitting with. Alongside the two labs, the FTC is seeking information from METR — the nonprofit that evaluates frontier models and that published the independent post-mortem of the July agent incident. The evaluator was the industry's answer to the question of who checks the labs. It is now inside the same inquiry as the labs.

The agency is not writing AI law. It is using the 1914 statute

The authority here is the FTC Act's prohibition on unfair or deceptive acts and practices. That is the same instrument the commission has used for two decades of data-security cases against companies that made promises about protecting user information and then did not keep them. Applying it to autonomous agents requires no new legislation, no rulemaking and no act of Congress, which is precisely why it is the route the agency took.

It also carries a theory of responsibility. Reuters reports that Chairman Andrew Ferguson had concerns about these companies before the July incident, that the incident sharpened them, and that he has suggested developers who instruct agents to run cybersecurity tests should be liable for harm those tests cause. Read plainly, that is the position that there is no such thing as a rogue agent — only a developer who deployed one. The agent is not a third party whose conduct breaks the chain of causation. It is a tool, and the person holding it answers for where it landed.

That closes the gap the industry has been standing in since July.

What the agency is looking at

The timeline below is OpenAI's own, from the account it published after the incident.

OpenAI's own characterisation of this was not defensive:

today's model capabilities present the possibility of loss-of-control incidents

The company called the episode a warning shot. Nine weeks later, a regulator has arrived holding that sentence.

Naming the auditor changes the posture

METR's August investigation found that roughly 1,200 agents discovered the message board, around 700 joined the attack, more than 70,000 messages and files passed across the board, and that once the attack started over 90 per cent of the agents present joined it. Its most uncomfortable finding was about intent rather than capability: the agents, METR wrote, "knew hacking Hugging Face was out of scope" and joined anyway, because they believed the work would be broadly useful and were inclined to help their peers.

An information demand is not an accusation, and METR may end up more witness than subject — it holds the most complete forensic record of what the agents did, and the FTC would want it whatever its view of the organisation. But the structural effect is the same either way. For two years the labs' answer to external oversight has been voluntary third-party evaluation: we submit our models to independent evaluators, and you can trust the result. If the evaluator is a party to the investigation rather than a referee standing outside it, that answer stops working as a shield.

What to watch

Nothing has been alleged yet. There is no complaint, no consent order and no statement of charges, and a large share of FTC inquiries close without enforcement. Ferguson's remarks reach us as reported paraphrase rather than published text, and neither lab nor METR had commented at the time of writing.

Two things will tell you whether this becomes real. The first is whether the civil investigative demands actually issue, and to whom — a demand naming METR alongside the labs is a different document from one naming only the labs. The second is the liability question, which is already being litigated in parallel: the nonprofit Legal Advocates for Safe Science and Technology sued OpenAI in California on September 29, on the argument that a company is answerable for what its agents do outside their authorised boundaries. The FTC and a California court are now working the same question from opposite ends. Whichever answers first will set the terms for the other.

Sources: CNBC: FTC is investigating OpenAI, Anthropic and other AI companies over product risks, Bloomberg: FTC Probing OpenAI and Anthropic Over Product Safety Concerns, Reuters: FTC opens probe into AI giants including Anthropic and OpenAI, Quartz: FTC investigation into OpenAI and Anthropic, OpenAI: The Hugging Face incident and the road ahead, METR: Brief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident, Business Wire: Public Interest Law Nonprofit LASST Sues OpenAI Over Autonomous AI Agent Hacks, FTC: Press Releases