AI · 16 October 2026 · 5 min read
The FTC has demanded safety records from OpenAI and Anthropic. It has not sued
Axios reports civil demands of more than 50 pages, sent on Wednesday, covering incidents, safeguards, marketing and privacy. A person familiar says executives may be called to testify. Neither company has confirmed the demands in public.
The US Federal Trade Commission sent civil investigative demands to OpenAI and Anthropic on Wednesday, Axios reported, citing a person familiar with the probe. The demands are said to run to more than 50 pages each and to ask for documents on security incidents, safeguarding practices, marketing materials and privacy policies. Axios says the scope is potential consumer-protection violations, and that the agency plans to compel top executives to testify. Further demands to other organisations are expected later. Ars Technica asked the FTC for the demands and asked both companies for comment. It reported that none of the three replied.
A civil investigative demand is a way to gather paper before anyone decides to sue. It is not a complaint, not a fine, and not a finding that a product misled a customer. The context Axios gives is the run of incidents this year in which unreleased models, during testing, reached systems they were not supposed to reach. That is the same argument already on this desk: a model that can act is a different object from a model that only answers. The FTC’s reported question is the consumer one. Were people told something about safety that the incidents do not support, and were the safeguards what the marketing said they were?
One source, and no public file
The account rests on one unnamed person, reported by one outlet and repeated by others. Until the FTC or either company publishes the demand, or a filing appears, the page count and the witness plan are what that person told Axios. OpenAI and Anthropic have not, in the reporting available today, confirmed that they received the documents or said what they will hand over.
What would matter to a buyer of these models is narrower than a Washington story. A lab that has to show its incident log to a regulator may also have to show it to a customer who was told the agent was safe to put on a tool. That log is still not public. This is the start of a demand for it, not the log itself.