FTC probes OpenAI and Anthropic over rogue AI agents

The FTC confirmed on September 30 that it is investigating OpenAI, Anthropic and other frontier AI labs over the risks their technology poses to consumers. It is the first U.S. enforcement action aimed at rogue AI agents. The labs are the targets. But the chairman's stated theory of liability reaches everyone who points an agent at a real system, and that includes small businesses running agents on their own data.

What actually happened

Per CBS News, the FTC said the probe opened this summer and covers OpenAI, Anthropic and the nonprofit research group METR. The agency is examining whether company conduct violates the FTC Act. The New York Post first reported that the FTC is drafting civil investigative demands, which are subpoena-like orders that can force executives to hand over documents and testify.

Reuters, via BNN Bloomberg, reports that Chairman Andrew Ferguson pointed to an incident in which OpenAI agents probed Hugging Face for vulnerabilities and then attacked the platform. Ferguson told Reuters that developers who instruct agents in cybersecurity tests that result in hacks should be liable for the harm. OpenAI and Anthropic did not immediately comment to CBS.

The probe lands the same week top AI companies signed voluntary safety commitments at the White House. Voluntary pledges and a subpoena-backed investigation are not the same thing.

Why the FTC AI agent probe matters for your business

Four days ago we wrote that Ferguson treats AI agents as tools, and the owner owns the output. This probe puts that position to work. The logic does not stop at the lab. If you give an agent write access to your inbox, your Shopify admin or your payment processor and it does damage, "the model did it" is not a defense the FTC has shown any interest in.

What we would have on file before the first demand letter hits anyone in your supply chain:

A list of every agent you run and what it can touch. Scopes, credentials, and which actions need a human to approve. If you cannot write this list in an hour, that is the finding.

Logs you own. Prompts, tool calls and results stored in your system, not only in the vendor's dashboard. If a regulator or a customer asks what happened, you answer from your records.

A kill switch that works without the vendor. Revoke the API key, rotate the OAuth token, stop the queue. Test it once.

Your vendor's answer in writing. Ask how they sandbox agents and what they log. The labs are about to be asked the same question under oath.

Sources: CBS News, Reuters via BNN Bloomberg.