This Meta Muse incident is a scary warning about trusting AI agents

It’s only been a few days since Meta launched Muse, and we’re already getting a sense of how much Meta wants its new AI agent to do. But the more an AI can do for you, the more you also have to trust it with. One early Muse experience shows why that can get uncomfortable very quickly.

Threads user Matt Robb shared what happened after using Muse to handle a Facebook Marketplace listing. According to screenshots he posted, the AI negotiated with a buyer, agreed to a price, shared his address, and arranged a pickup — all without him realizing someone had actually been invited over.

The buyer eventually showed up at his building and waited, and Robb said he only learned about the situation afterward, when Muse told him what had happened. He then instructed Muse not to agree to any future pickups without checking with him first.

That’s where this goes from an AI mistake to something much more serious. Letting an AI deal with the repetitive side of Facebook Marketplace sounds useful. It could answer endless availability messages or negotiate within a price range you’ve already set. But sharing your address and deciding when someone should come to your home is a very different responsibility.

Another user who shared Robb’s experience on X called the situation “dangerous and creepy” and said they deleted Muse after seeing it. The concern is easy to understand when an AI action can result in a stranger physically turning up at someone’s home. But one big question remains: how did this happen in the first place?

David Singleton of Meta Superintelligence Labs responded to Robb on X and said he had reached out privately to investigate the incident. Singleton also said Meta had looked into similar reports in the past and had consistently found that Muse followed direct instructions and correctly asked users for permission.

It’s a careful response, and importantly, it doesn’t tell us what happened in Robb’s particular case. Singleton said Meta is looking into it, so it would be unfair to conclude that Robb mistakenly approved everything Muse did.

At the same time, the screenshots and Robb’s account alone aren’t enough to show that Muse ignored or overstepped a permission. Muse could ultimately have done exactly what it was authorized to do. It could also turn out that something went wrong with how those instructions or permissions were interpreted or carried out. Until Meta finishes investigating, both possibilities are worth keeping open.

And that uncertainty is exactly why I’d be extremely careful about handing something like this over to an AI agent in the first place. AI agents are exciting precisely because they can go beyond answering questions and start doing things for us.

This isn’t unique to Muse, either. Gemini, ChatGPT, Claude, and other AI assistants can increasingly connect with our apps and services and take action for us. That can save a lot of time, but it also means we have to think much harder about which decisions we’re comfortable handing over. Personally, arranging an in-person Marketplace pickup would be well beyond that line for me.

We still don’t know whether Muse or a permission granted to it was responsible for what happened in Robb’s case. Meta’s investigation may eventually give us a clearer answer. But the fact that multiple plausible scenarios could have led to this is enough of a warning on its own. AI agents will ask for more trust as they become more capable, but incidents like this remind us we don’t necessarily have to give it to them.