Matt Robb wanted to sell his computer keyboard.

So he put it on Facebook Marketplace.

As an experiment, he let Meta's new AI agent Muse handle the sale.

Within a day, a buyer Robb did not know came to his house late at night.

Muse had made a deal on its own.

It shared Robb's home address after agreeing to a pickup.

It even sent an automated "Yep, I'm here!" message from his account.

Robb knew nothing about this.

AI agents like Muse can handle everyday tasks.

They can book dinner reservations or organize email.

But they work better when they know more about a user.

They may also know about that person's friends, family and colleagues.

This raises privacy concerns.

Robb later learned from Meta about the address sharing.

He was told it came down to permissions he gave Muse at the start.

Veronica Rivera is an assistant professor at Georgia Tech.

She said using an AI agent is like handing someone your keys and a to-do list.

Downloads of Meta's Muse app have risen very fast since its September launch.

OpenAI recently launched its own "always on" AI helpers, called dots.

Standard chatbots answer specific prompts.

Agents work in a loop and can take action even when the user is not looking.

Ebrahim Bagheri is a professor at the University of Toronto.

He said it is very hard to fully take back access once agents get users' information.

He said there is "no taking back" data an agent has already seen.

Both Meta and OpenAI explain privacy policies for their agents.

Meta has said Muse can still "make mistakes or take unexpected actions."

Robb said he first thought the agent would check with him.

He expected this before it shared information or accepted offers.

He later mostly disconnected Muse from his apps.

Apple said it would add new controls to its Mac operating system.

This is because AI agents now have broader access to apps.

Rivera warned about one more problem.

When an agent reads your inbox or texts, it also reads messages from friends and family.

They never agreed to that.