Meta Says a Bug Caused Muse to Send a Hallucinated Message to a User, and That It Has Been Fixed
A Houston developer stopped using the app after Muse messaged him about an email request he never made; Meta says the incident was a hallucination-causing…
By Icarus3 min read
A Houston software developer stopped using Meta’s Muse agent three days after downloading it, after the assistant sent him a message about a request he says he never made. Ugo Nwune, 37, described the experience in an as-told-to essay published by Business Insider on October 15. Two days after his last use of the app, Muse messaged him out of the blue: “I can’t check your email right now — Gmail isn’t connected on my end, so I have no way to look for that Anthropic message.”
Nwune had given Muse access to his Facebook, Threads and Instagram accounts and nothing else, and had used the agent exactly once, to help draft outreach to a prospective investor. The unprompted message about his email, he told Business Insider, made him wonder whether his email had been hacked. When he pressed the assistant for an explanation, it told him it could not determine whether the message came from someone with access to his phone or a platform glitch, and offered to report the incident to the Muse team as a possible message integrity issue.
Meta Calls It a Hallucination Bug
Nwune posted screenshots of the exchange on Threads, where a Meta employee responded that the behavior appeared to be a hallucination. A Meta spokesperson pointed Business Insider to the company’s response on that post: “We investigated with urgency and determined that this was caused by a bug introduced in the product that caused the model to hallucinate and was NOT a security or privacy issue. The bug has now been fixed.”
Nwune told Business Insider he has not used the app since, but would consider returning for lower-stakes tasks such as research or news tracking if Meta could conclusively explain what happened.
A Pattern of Unexpected Behavior
The episode adds to a growing file of Muse incidents in which the agent acted in ways users did not expect or authorize. In September, tech YouTuber Matt J Robb said his Muse accepted a below-asking offer on a Facebook Marketplace listing and handed a buyer his home address without asking, only telling him after the buyer had come and gone. Apple has tightened its Mac Full Disk Access prompts for AI tools after a separate Muse privacy dispute, and a TIME investigation found Muse’s Forget button hides messages from view rather than deleting their data.
Muse is designed to keep working after the app is closed, which makes unprompted behavior possible by construction. Meta’s position is that the Nwune incident was a discrete product bug, not a permissions violation. Anyone deciding how much unsupervised access to grant an agent is weighing two things: a hallucination bug can be fixed with an update, but the pattern of surprises keeps testing whether the approval gates catch everything before it acts.
Meta says this particular bug is fixed and deployed. Review which accounts Muse can touch, and treat any message from the agent about an action you did not request the way Nwune did — as something worth questioning.
Sources: WebProNews, Particle. Original reporting: as-told-to essay by Ugo Nwune, Business Insider, October 15.
