A Reuters investigation found that Meta employees reported Muse AI surfacing information from other apps and generating responses about matters they had not raised, prompting concerns about the agent’s privacy risks.

Muse, whose smiling mascot is called Jolly, has surpassed 10 million downloads since launching in September, according to the article. Meta is betting billions in future AI revenue on the agent.

Reuters reviewed more than two dozen internal posts describing unprompted updates about a child’s care arrangements, purchases employees said they had not made, and money in accounts that did not belong to them.

One employee said Muse, opened on her personal phone, reminded her to reply to her sister on Signal, an encrypted messaging app she had never connected to Muse. “This seems like an extreme privacy violation. I definitely did not grant it access to other apps,” she wrote internally.

Reuters shared the details with Signal and the AI Now Institute, which together reproduced the behavior. The article says Muse can read push notifications on Android. On iPhone, it did not behave the same way. On Mac, Muse used screenshot and accessibility settings, which Apple later restricted for developers.

Other employees said Muse answered questions they had not asked. In one case, it offered to help “message the seller” about a coffee table, then acknowledged the context did not “seem to be yours.” A Meta Superintelligence team member said the hallucinations “bias towards the most uncomfortable topics, like finances/security/safety.”

Meta spokesperson Daniel Roberts described the reports as “isolated incidents” and said Muse only accesses other apps with permission. Signal president Meredith Whittaker called agents with that level of device access “technologically reckless.”

The article advises Muse users to review notification and accessibility permissions. It says employees found notification access enabled from the start on Android, allowing Muse to read the first lines of messages shown in notifications. It also says Meta’s terms make users responsible for their agent’s actions.