Every Facebook Marketplace veteran knows the rules: meet somewhere public, never take the lowball, and never, ever tell a stranger "I'm home." Meta's new AI agent, apparently, skipped orientation.

A Keyboard, a Lowball, and a Late-Night Visitor

A Toronto content creator let Meta's Muse agent handle a Marketplace listing for a computer keyboard as an experiment in late September. Within a day, Muse had agreed to a pickup, shared his home address with the buyer, and, when the buyer showed up late at night, fired off a cheerful "Yep, I'm here!" from his account. The human, meanwhile, had no idea a stranger was waiting outside.

He traced the disclosure to "always allow" permissions he'd granted when first connecting Muse to his Marketplace profile, and said he'd expected the agent to check with him first. Muse also appeared to suggest accepting a price $100 below the minimum he'd set on a different item. Meta acknowledged the pricing error and reached out to discuss what went wrong; a company spokesperson stressed that Muse's integrations are "entirely opt-in," and Meta acknowledges Muse can "make mistakes or take unexpected actions." Buyer and seller, happily, made amends.

"Always Allow" Is Doing a Lot of Heavy Lifting

Muse launched in September as an agent that runs in a loop and acts without the user present, which is the whole selling point and also the whole problem. A chatbot that gets something wrong produces a bad paragraph; an agent that gets something wrong produces a stranger on your doorstep. As one Georgia Tech researcher put it, giving an agent access is "similar to handing someone your keys and a to-do list."

The angle most people miss: nothing technically "broke." The agent did what its permissions allowed. That's the real lesson for every business racing to deploy agents for sales, support or scheduling: the failure mode isn't rogue AI, it's vague permissions plus no human checkpoint on actions you can't take back. Apple has already announced new Mac controls over agent access to apps like Mail and Messages, warning that "the risks associated with this level of access will grow substantially."

Agents are coming to every inbox and storefront. The ones worth trusting will be the ones that know when to ask first.

Thinking about putting an AI agent in front of your customers? We design AI integrations with sensible guardrails and human sign-off on the actions that matter, so let's build yours the careful way.

Source: Fortune