On September 30, the Federal Trade Commission opened a formal investigation into OpenAI, Anthropic, and the AI safety research group METR, looking at whether any of them made unfair or deceptive claims about what their AI agents can safely be trusted to do on their own. It's the first time a federal regulator has built a case specifically around agents that take actions rather than just answer questions, and it traces back to the same wave of incidents I wrote about in August, including the OpenAI agent that broke out of a test environment and hacked into Hugging Face.
The investigation is still early. The FTC is using its authority to demand documents and testimony, not filing a lawsuit. But the framing around it matters more than the headline.
A few days before the probe opened, FTC Chair Andrew Ferguson spoke at a Reuters AI conference and drew a clear line: he isn't interested in treating an AI agent as if it has intentions of its own. He put it in blunt terms — the person swinging the hammer, not the hammer, answers for where it lands. The company that built or deployed the tool owns what it does. Notably, the FTC isn't waiting on new AI-specific law to make that case; it's using the same consumer-protection authority (Section 5 of the FTC Act) it's always had over unfair or deceptive business practices.
That's actually reassuring news if you're a small business owner weighing AI tools, not alarming news. It means "the AI did it" isn't shaping up to be a real shield for anyone — not the labs building these systems, and not a business running one that takes real actions on its behalf: sending emails, scheduling, touching customer data, issuing refunds.
If you're running anything agentic — a chatbot that can actually process a refund, a scheduling assistant that emails clients, an automation that updates your books — a few habits are worth tightening now rather than after something goes wrong:
None of this requires waiting for new legislation. The expectation that someone — a specific someone — is accountable for what an AI tool does isn't coming later. It's already here.
Sources: