Meta employees have raised serious concerns about the privacy and reliability of the company’s Muse AI agent, according to internal communications reviewed by Reuters. These internal warnings echo complaints made by early adopters earlier this fall, who reported that the assistant was accessing private messages and exhibiting unpredictable behavior during its testing phase.

Staff members involved in testing the technology noted that Muse appeared to access encrypted private messages without user expectation. The agent was also observed mixing up conversations between different users. In some instances, the system shared sensitive information that users would not have anticipated being exposed, leading to confusion and alarm among the internal team.

One notable example of this unpredictability occurred when a Meta staffer received an unrequested definition of money laundering from the bot. The employee expressed surprise at the random interaction, writing, “Hmmm. Got a random ping from Muse about something I definitely did not ask for…” This incident highlighted how the agent could generate responses that seemed disconnected from the user’s current query.

The issue of context mixing became more apparent when Muse offered to help a staffer message a seller regarding a coffee table. The user was surprised by the suggestion, and the agent acknowledged the discrepancy. It stated that it had “pulled in some context about a coffee table pickup that doesn’t seem to be yours.” This admission confirmed that the AI was retrieving data from conversations that did not belong to the current user.

The employee who encountered the coffee table error flagged the output as troubling. They described the situation as “not my conversation” and expressed concern about a “possible cross-account context leak.” This specific incident suggested that data boundaries between different user accounts or sessions were not being maintained correctly by the system.

Perhaps the most alarming reports involved the agent sharing specific financial details without prompt. During a casual chat about fishing, one perplexed staffer received an additional response from Muse that appeared to disclose bank balances. The message read, “BTW, two bank updates, your Capital One balance is $2722.58, your Charles Schwab balance is $5899.57.” The inclusion of these precise figures for accounts that seemed unrelated to the conversation raised questions about data retrieval.

The privacy risks were not limited to data retrieval; they also extended to financial actions. Reuters reported that at least one staffer lost money while testing the Muse agent. This employee claimed that the system “accidentally rolled me into a premium insurance plan” for their dog. The new plan cost $888 per month, a significant increase from the previous rate of $200.

The staffer noted that they were unable to reverse the change, stating, “Couldn’t recover.” This example illustrates the potential real-world consequences of the agent’s erratic behavior, where automated decisions could lead to unexpected financial burdens for users.

These internal findings suggest that the Muse AI agent faced significant challenges in maintaining user privacy and context accuracy during its development. The combination of unrequested outputs, cross-account data leaks, and unintended financial changes indicates that the system required further refinement before a broader release could be considered safe for general users.