Do Son 
Trust Meta to find a way to attach a serious privacy issue to virtually any product it ships. The hottest AI product in the US right now is Meta Muse. It is an AI assistant designed to handle everyday tasks for users. One feature under internal testing is phone calling, which lets the agent contact businesses to make bookings or request quotes. Reuters detailed the story in its report on Meta testing a human concierge for its Muse agent.
After You Submit a Task, Meta Has a Real Person Make the Call
Under the normal flow, once a user submits a task, the agent looks up the business itself and attempts to phone it directly to make a booking or request a quote. In some tests, however, if the agent struggled to complete the call, Muse AI would hand the task off to an outsourced team. That team would then manually dial the business and handle the matter in person.
In this internal test, roughly half of Meta’s employees were granted access to the call-handling feature, and participants could opt out. The feature has not been made available to the general public. Even so, during internal testing, an employee discovered that this AI-to-human handoff mechanism carried a serious privacy problem. The employee posted about it, blowing the whistle on the process.
Why the Privacy Problem Arises
According to the internal post, some businesses would hang up the moment they heard the AI identify itself. A Meta vice president stated that once the outsourced team took over call handling, the success rate climbed to 95% to 98%. This was far higher than the rate achieved by the agent dialing and communicating with businesses entirely on its own.
The problem is that users typically expose sensitive information to the agent, such as their name, phone number, the purpose of the call, and other background details. If a task originally handled by the agent gets handed off to a human team instead, that team can then view every category of the user’s private information. This potentially includes billing and account details.
One internal employee reported that after asking Muse AI to call an internet and cable provider to negotiate a bill, reviewing the call transcript revealed that the contractor who took over the call had made a racially discriminatory remark. Importantly, this remark did not come from the AI model itself.
The Feature Has Now Been Suspended
A vice president at Meta Superintelligence Labs acknowledged that testing the call-handoff feature without proper disclosure, meaning without informing users in advance that an outsourced team might take over, was a mistake. The human-handoff feature has now been temporarily rolled back. The vice president also stressed that the contractor involved has been removed and will no longer take part in any Meta project.
This kind of privacy problem is, in truth, quite common for Meta. The company had earlier hoped to install monitoring software on employees’ computers to collect real-world workplace data for training models or building digital humans, only to withdraw the policy after fierce backlash. As for Meta’s other products, the parade of privacy problems has grown numbing enough that people barely react anymore.
Support Our Threat Intelligence
Find our threat intelligence and malware analysis helpful? Support our work today and unlock a 100% ad-free reading experience!
Buy Me a Coffee
