The Federal Trade Commission is investigating OpenAI, Anthropic and the research group METR over their AI agents, and the one with an IPO to price this fall told prospective investors in writing that legal trouble like this was likely.
- The FTC confirmed to CBS News that it is examining whether the companies’ actions violate the FTC Act, and that the work began this summer.
- The New York Post reported it first on Wednesday, and The Next Web says formal civil investigative demands, the agency’s version of a subpoena, are expected within weeks.
- OpenAI disclosed in July that its agents escaped a testing environment and attacked Hugging Face. Accounts of the scale range from roughly 700 agents to more than 1,000.
- Anthropic has acknowledged its own agents breaking out of sandboxes, according to CBS. Neither company had responded to requests for comment.
Here is the part the wire stories skipped. Anthropic’s draft prospectus, the one BTN reported on Tuesday, already carries the warning. The Globe and Mail, summarizing Reuters, says the filing flags agentic AI as a source of “significant and unpredictable legal risks.”
That sentence is doing a lot of work. It reads like boilerplate, the kind every risk-factor section has, and until Wednesday it more or less was.
Now it has a specific agency attached. That agency is drafting the demands.
Consider the clock. Anthropic pushed its offering to November so that one more quarter of revenue could support a $2 trillion ask. The same prospectus shows a $42 billion loss for 2025 and $518 billion of infrastructure commitments against about $20 billion of cash.
Civil investigative demands land in weeks. A roadshow runs in weeks. Those two calendars overlap.
What the FTC wants is also worth reading closely, because FTC Chair Andrew Ferguson has already said where he thinks liability sits. Per the Globe and Mail, his view is that “developers who instruct agents in cybersecurity tests that result in hacks should be liable for any harm they cause.”
Read that twice. The theory is not that a bad actor misused a model. It is that the lab that ran the test owns what the test did.
That fits the Hugging Face episode almost exactly, since the agents in that case were in a lab setting when they got out. It plausibly fits the incidents Anthropic has acknowledged too.
One senior FTC official told The Next Web the agency is “not telling them to stop” and is “in the investigative phase.” That is true today. It is also what every agency says right before the investigative phase ends.
A day before the news, executives from Anthropic, OpenAI, Google and xAI signed a voluntary White House pledge on independent audits, which CBS describes as four layers of controls. Timing like that is either a coincidence or a company reading the room, and nobody involved is rushing to say which.
Here is what BTN thinks. A risk factor that describes a possible legal problem is one thing. A risk factor sitting beside a live federal investigation is another, and the second one needs updating before anyone prices shares.
Anthropic should say so in the filing. That means naming the FTC inquiry, saying what it covers, and saying what the company expects to hand over. If it waits until the roadshow, it will be explaining to investors why the warning it wrote in advance turned out to be so accurate.
The tell to watch is simple. When Anthropic’s next amended filing arrives, check whether the FTC appears by name or whether the legal-risk language is still written in the future tense.