Meta's Muse agent messaged a user out of nowhere; Meta calls it a hallucination bug and says it's fixed

A software engineer's experience with Meta's Muse agent has become the latest data point in the product's reliability debate. In an as-told-to essay for Business Insider, Ugo Nwune said he wanted to try Meta's agent because an AI from a trillion-dollar company felt safer than the wild west of open-source agents. Futurism reported on the account on October 18.

When he set up Muse, Nwune stressed that he granted it no permissions on his device beyond access to his Facebook, Threads and Instagram accounts. He tried it, then left it untouched for a few days, at which point the agent began to act strange and sent him a message he had not prompted.

The message read: 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 was confused, because he had not asked the agent to check his mailbox and had not messaged it in two days. He briefly thought his email might have been hacked.

When he pressed Muse for an explanation, it offered another puzzling line: it could not tell from its side whether someone with his phone or a platform glitch was responsible, and it would not pretend otherwise. Nwune reported the exchange to Meta, where an employee told him it appeared to be a hallucination. He did not find that satisfying.

I haven't used my Muse agent since, Nwune said, adding that what caught him off guard most was that he had done nothing to prompt it: he had not opened the app for two days and it acted on its own. In a statement to Business Insider, Meta insisted the AI simply did not know what it was talking about and had not violated any permissions.

Meta's full statement said the company investigated with urgency and determined that a bug introduced in the product caused the model to hallucinate, that it was not a security or privacy issue, and that the bug has now been fixed. That places the blame on a repaired product defect without explaining why the model invented a specific Anthropic message in the first place.

The incident has been amplified by Muse's record so far. Meta encourages users to connect their email and bank accounts to the agent; Wired reported that it keeps detailed files on users' friends and family; one man said his Muse agent gave his address to strangers on Facebook Marketplace and sent them to his home; and an Inc Magazine journalist complained that Muse surfaced details from text messages he had not given it access to.

According to a New York Times scoop, Meta engineers worried about the agent's bugginess and frequent safety errors before launch, and the release slipped while they worked through them, until CEO Mark Zuckerberg pushed it out after seeing the success of Instinct, a rival agent from a small startup. Set against that backdrop, the problem looks less like a single bug than a product shipped before its reliability was settled.

The difficulty for users is that when an agent initiates messages and holds account access, a software defect and a security incident look identical on the screen, and only the platform can explain the difference after the fact. Fixing the bug is one task; giving users an auditable record of what the agent did and why is another. The next signal to watch is whether Meta publishes more detail on the failure and makes visible changes to permissions and activity logs.