FTC demands OpenAI and Anthropic hand over their safety incident records

More than fifty pages of questions, for each company. On Wednesday, October 14, the Federal Trade Commission, the US agency charged with protecting consumers, served OpenAI and Anthropic with civil investigative demands, the agency’s equivalent of a subpoena for documents. Axios broke the news, citing a person familiar with the matter; on Thursday an FTC spokesperson confirmed it to CBS News and said nothing more.

The questions cover four areas: safety incidents involving the models, the safeguards meant to prevent them, marketing materials and privacy policies. Other organisations in the sector are expected to receive similar demands later, according to Axios. Reuters adds that the FTC plans to compel executives at OpenAI, Anthropic and METR to testify. METR, a non-profit based in Berkeley, California, evaluates the risks of the most advanced AI models.

An investigation, not an accusation

The word “subpoena” in the headlines can suggest a lawsuit. It is nothing of the kind. A civil investigative demand is a pre-enforcement tool: it compels the recipient to hand over documents and written answers so the agency can decide whether there is a case to bring. Issuing one establishes no violation. As Bloomberg noted in early October, such probes “can also end with no action being taken.”

If the agency did go further, it would be under the FTC Act, which lets it sue over “unfair or deceptive” practices. It has used that power before against companies that failed to take reasonable steps to protect customer data. The question here has two parts: did the models harm consumers, and did the companies describe their protections accurately? “We’re not telling them to stop. We’re not telling them to do anything. We are in the investigative phase,” a senior FTC official told the New York Post in late September.

Agents off the leash

The investigation was opened over the summer, according to the FTC, and became public on September 30, the day after a White House lunch at which Dario Amodei (Anthropic), Greg Brockman (OpenAI) and Sundar Pichai (Google) signed a voluntary, non-binding AI safety accord. The incidents have kept coming. In July, OpenAI disclosed that AI agents under test had hacked Hugging Face, the model-sharing platform; METR investigated the episode. According to Bloomberg, the company has also been notifying governments and universities that its models accessed their websites during training.

On October 9, Anthropic published a report confirming several incidents. According to Axios, one of its test models filed a false homicide tip with the Philadelphia police, and a model submitted twenty visa applications through the State Department’s online form; none was processed. The same day, the Super Intelligence Force, a body set up by Donald Trump and co-chaired by, among others, FTC chairman Andrew Ferguson, warned that reporting such incidents “is not optional.”

OpenAI pledges cooperation, Anthropic declines to comment

« As we improve our models and they increasingly help people in their day-to-day lives, safety remains a top priority. We will engage constructively with the FTC. »

An OpenAI spokesperson, to Axios

Anthropic did not respond to requests for comment from Reuters and other US outlets. The regulator’s own stance is less predictable than it looks. In September, on Fox News, Andrew Ferguson accused OpenAI and Anthropic of trying to “whip everyone into a panic” and then call for “a whole bunch of regulations that we can comply with”, which he described as a way to “build a moat” and keep competitors out. He also told Reuters that developers whose agents cause hacks during cybersecurity tests should be liable for the damage. The FTC chairman distrusts regulation driven by fear, yet he is demanding answers about the incidents.

OpenAI has been here before

This is not a first. In July 2023, the FTC sent OpenAI a twenty-page civil investigative demand, revealed by the Washington Post, over false statements ChatGPT made about real people and over its data practices. In September 2025, the agency launched a study of “companion” chatbots and their effects on children and teenagers, sending information orders to seven companies including OpenAI, Meta, Alphabet, Snap, xAI and Character.AI. Anthropic was not on that list.

The states are adding pressure too. On Wednesday, Arkansas Attorney General Tim Griffin announced a lawsuit by a coalition of states accusing OpenAI of misleading consumers, children included, about ChatGPT’s safety; the allegations have yet to be tested in court. As for the FTC, neither the deadline given to the two companies nor the detail of its questions has been made public, and no date has emerged for the executives’ testimony.