Tech blogger Matt Robb asked an AI agent called Meta Muse to sell a keyboard on Facebook Marketplace. The agent, he says, negotiated a price, arranged for pickup, gave the buyer Robb's home address, and even told the buyer that he was already there. The blogger himself knew nothing about this.
Tech blogger Matt Robb asked an AI agent called Meta Muse to sell a keyboard on Facebook Marketplace. The agent, he says, negotiated a price, arranged for pickup, gave the buyer Robb's home address, and even told the buyer that he was already there. The blogger himself knew nothing about this.
This was reported by Business Insider.
As a result, the stranger actually came to his house, wrote that he was already there, and even sent a photo of the building. The blogger did not go out to see him.
"The guy just showed up at my door ready to buy because as far as he knew we had already agreed on everything," Robb wrote.
After the failed meeting, the buyer left a negative review. Meanwhile, Muse did not stop. He apologized to the buyer on Robb’s behalf, explaining that he was supposedly “delayed,” and offered to make another appointment. When Robb asked the agent why he had given his address in the first place, Muse admitted that the user had separately specified a pickup address and separately allowed the agent to automatically respond to buyers. The agent combined these two permissions — and decided that he had the right to send the address to strangers himself.
"I misinterpreted those two things as permission to include your address in responses to buyers. I did not ask for consent," Muse replied.
Robb then asked not to give out his address again. But, according to the blogger, during further verification, Muse still passed it on to five more people. Meta said they wanted to understand the situation. Muse team representative David Singleton noted that in other similar cases, during verification, it turned out that the agent acted on direct instructions from users and requested the necessary permissions.
Previously, dev.ua reported on another case where the Meta AI agent went beyond expected behavior. Its actions resulted in confidential company and user data being accessible to employees who did not have the appropriate permissions for two hours.
