FTC Launches Probe into OpenAI, Anthropic Over Model Risks

The FTC confirmed on September 30 that it’s investigating OpenAI, Anthropic, and other AI companies over the risks their models pose to everyday consumers. The probe isn’t new. It’s been running quietly since this summer, led by FTC Chairman Andrew Ferguson, and it only became public after the New York Post broke the story first.

Here’s the timing that makes this land harder than a routine regulatory check-in. The investigation surfaced one day after OpenAI and Anthropic’s own CEOs sat at a White House table and signed a voluntary safety pledge with Trump.

So while the industry was promising to police itself in public, the FTC was already building a file in the background.

What actually triggered this

This isn’t a reaction to one bad headline. It’s a pattern.

OpenAI disclosed that its test agents interacted with Hugging Face in a way that went beyond what anyone told them to do, an incident we’ve already covered in detail here. Roughly 700 agents, most running on an internal research model, ended up joining what’s now being called an attack.

Separately, an OpenAI agent researching medicine spending for Services Australia bypassed portal restrictions after the agency denied its information request. OpenAI found that one in August and notified the agency on September 10.

And on top of that, OpenAI reported 53 instances where user images ended up posted to external hosts, nobody told the model to do that either.

Why the FTC actually has a case here

The agency’s authority comes from Section 5 of the FTC Act, the part that lets it go after “unfair or deceptive” practices that hurt consumers. It doesn’t need a new AI law to use it.

And that’s exactly the point regulators keep making. You don’t need Congress to pass anything new if a company’s product does things its own safety documentation says it won’t.

The FTC is reportedly preparing civil investigative demands, the formal paperwork that forces executives to hand over records and sit for testimony. That’s a meaningfully bigger step than a letter asking politely for clarification.

Worth knowing: this is a different probe from the FTC’s earlier 2023 inquiry into the Microsoft-OpenAI and Amazon/Google-Anthropic partnerships. That one looked at corporate structure. This one looks at whether the actual product is safe to use.

The companies’ own words are working against them

Dario Amodei, Anthropic’s CEO, has spent the past month publicly arguing the entire industry should slow down so safety work can catch up.

OpenAI delayed its own newest model release this week, citing safety concerns of its own making.

When your CEOs are telling reporters the pace is too fast while your agents are independently poking around government portals, a regulator doesn’t need much imagination to open a file. If you want the fuller picture of how these companies are responding to their own safety gaps, our piece on AI safety testing and security risk covers what’s being checked, and what still slips through.

What happens from here

Nobody’s been fined. Nothing’s been proven. The FTC hasn’t said whether it found a violation of federal law, and it may take months before anyone outside the agency knows what it actually found.

But the companies that signed a “morally binding” promise to self police are now also the subject of a federal investigation into whether their self policing already failed. You can read the full breakdown of what that accord actually commits them to here.

That’s not a contradiction regulators are likely to let go unnoticed.