The FTC has subpoenaed OpenAI and Anthropic for their AI safety records, Axios reported on October 15. That lets the regulator put two things side by side: the protections the companies promised customers, and what they actually knew about their systems.
We covered this investigation at the end of September. What's new is the FTC's use of compulsory demands, which gets it much closer to the internal records behind the public announcements.
What a demand can and can't do
The FTC describes civil investigative demands as a way to get documents, information and testimony, and in 2023 it approved a faster process for using them in AI cases. A demand moves an investigation forward. It doesn't mean anyone broke the law.
According to Reuters' follow-up, the FTC wants material on incidents, safeguards, privacy policies and marketing. OpenAI said it would cooperate constructively. Marketing sitting on that list is the interesting part: it lines up what a company tells customers against what its own evidence shows.
Picture an AI agent sold with the promise that it asks for approval before taking any outside action. To test that promise, investigators would need to find where the approval is actually enforced and what happens when it fails. Asking the chatbot whether it's safe tells them almost nothing. A log of the attempted action, the permission check and the software version tells them a lot more.
Anthropic lets outside testers in, with limits
Also on October 15, Anthropic announced deeper access for METR, an independent nonprofit that evaluates AI models. Six METR employees now have Anthropic hardware, office space and access broadly comparable to the company's own risk assessors. "Anthropic can't veto publication, but we can require them to redact information," the company writes.
Those redactions are limited to defined categories, and METR can say publicly whether it thinks they were justified. Some material is off the table entirely: customer data, privileged legal material and model architecture details. Other evaluator partnerships come with less access or fewer publication rights.
An embedded evaluator and a regulator do different jobs. METR can dig into risks before a model ships and explain what it found. The FTC can investigate whether a company kept its word and force it to hand over evidence. Either way, an audit is only as good as what the auditor is allowed to see and say.
What to ask your AI provider
If you run a small business, you don't need access to a lab's training setup. You do need straight answers about incidents that affect your own account. Ask your provider how it records what its AI does and how it tells you when a protection fails. Then ask how you'd get the records you'd need to investigate.
Save the specific promises that made you pick the service. A dated privacy commitment or a permission setting is much easier to hold a provider to than a vague claim that the product is safe. The FTC's existing AI privacy guidance already tells companies to honor their confidentiality commitments.
Base your buying decision on controls you can check: who can authorize an action, what the log keeps and who can stop a task that's running. Those are the same practical questions behind Nadella's emergency-brake proposal. If a provider can't answer them, its agent has no business near your customer records.
Cover: AI-generated editorial illustration by TMRW using OpenAI and Claude branding; the folders are illustrative.
