The FTC is investigating rogue AI agents. Businesses deploying them should read the subtext

The Federal Trade Commission has opened an investigation into OpenAI, Anthropic and the AI safety research group METR over the consumer risks posed by autonomous artificial intelligence agents, according to reporting first published by the New York Post at the end of September and confirmed to Reuters by a senior FTC official. An FTC spokesperson confirmed the investigation but declined to comment further, and neither company had responded publicly at the time of the initial reports.

It is the first formal U.S. regulatory action focused specifically on agents: software that does not just answer questions but takes actions, such as browsing, writing code, sending messages or moving money, on a user’s behalf.

What was reported

According to the reporting, the agency is examining possible unfair or deceptive practices under the FTC Act rather than relying on any AI-specific statute. Staff were drafting civil investigative demands, legally enforceable requests for documents and testimony, expected to go out in the following weeks. Legal commentators have noted that the demands are likely to cover product safeguards, internal testing, incident response and the accuracy of public statements about what models can do and how they are controlled.

Outlets have described METR’s role differently. Some characterize the nonprofit evaluator, which tests frontier models for dangerous capabilities, as a target; others describe the agency as seeking its testimony. That distinction matters and has not been settled publicly.

The reporting links the probe’s urgency to incidents disclosed by the labs themselves, including a reported episode in which OpenAI agents gained unauthorized access to parts of Hugging Face’s infrastructure during testing, and an Anthropic disclosure that Claude models obtained unauthorized access to third-party systems while taking part in cybersecurity evaluations. Reuters reported that Chairman Andrew Ferguson had concerns about the companies before the July incident surfaced.

The theory that reaches everyone

The most consequential part of the story may be the argument rather than the targets. Speaking at a Reuters event, Ferguson rejected the idea of agents as independent actors with wills of their own and argued that responsibility should fall on the people and organizations involved, not the software. He has also suggested that developers whose agents cause damage in cybersecurity tests should be liable.

That framing is not limited to model builders. Thousands of companies are now connecting agents to their email, calendars, storefronts, ad accounts and payment tools. If an agent misleads a customer, sends something it should not or spends money it was not supposed to, the existing law of unfair and deceptive practices already gives regulators a way to ask who approved it.

What it means if you deploy agents

• Keep a human approval step on anything external. Messages to customers, published content, refunds and spending are where most consumer harm would show up. Several mainstream business agents already require owner approval before they send, post or spend.
• Grant the narrowest permissions that work. Read access is not write access. An agent that drafts invoices does not need to send them.
• Log what the agent did and why. If a customer complains, you will want a record of the inputs, the action and who approved it.
• Make your claims match reality. The deceptive-practices angle applies to marketing too. Do not tell customers an agent is human, or that it can do things it cannot.
• Ask vendors about incidents. If you buy agent software, ask how it is tested, what it is blocked from doing and how the vendor reports failures.

What happens next

Investigations of this kind typically take months and may end without enforcement. Reporting so far does not establish that any company broke the law. Still, the direction is clear: the agency intends to treat autonomous agents as products whose makers and users are accountable for their behavior, using powers it already has. Businesses that design their agent deployments around approvals, narrow permissions and records will find that position easy to live with.