Meta Muse Hallucination: The Gmail Message Meta Calls a Bug

The Meta Muse hallucination became public on October 15, 2026, when a Houston software developer described a message he never asked for. His agent brought up his Gmail, an account he had never connected to it. Meta investigated and blamed a product bug that made the model invent the exchange. According to the company, the agent never touched his data.

Ugo Nwune is a software developer in Houston. He described the episode in an as-told-to essay for Business Insider, published on October 15, 2026. He had picked Muse over open-source agents because he trusted an AI from a trillion dollar company.

Crucially, he had connected only three accounts: Facebook, Instagram and Threads. He had not messaged the agent in two days. Then Muse wrote to him unprompted about his inbox.

“I can’t check your email right now,” the agent told him. “Gmail isn’t connected on my end, so I have no way to look for that Anthropic message.” He had never mentioned an Anthropic message, and he had never linked Gmail at all.

Nwune reported the message. A Meta employee told him it looked like a hallucination, an answer he did not find satisfying. The company then looked at it more formally.

Meta says it investigated with urgency and traced the behaviour to a bug it had introduced, which caused the model to hallucinate. In its statement to Business Insider, reported by Futurism, Meta said the incident was not a security or privacy issue. Meta has since fixed the bug.

In other words, the agent did not reach into anything. It only talked as though it had.

That distinction did little to settle him. Because the message arrived on its own, his first thought was that someone else had got into his accounts. A machine volunteering details about an inbox tends to read that way.

So he pressed the agent for an explanation. Muse replied: “I can’t tell from my side whether it was someone with your phone or a platform glitch, and I won’t pretend otherwise.” The hedge did not help either.

As a result, Nwune cut off the agent’s access to his accounts and stopped using it. He had not prompted the message, he said, and it just did that on its own.

This was not the first time the Meta Muse agent stated something untrue with confidence. A hands-on review at Yahoo Finance on September 23, 2026 found the same pattern during ordinary errands.

When he asked about a casual restaurant, Muse supplied a phone number, then sent a different one. When he queried the change, it admitted it had made the first number up out of thin air and had not verified it. The same test also surfaced a venue that had closed years earlier.

Here is what reviewers and users have reported so far.

Muse is a consumer agent rather than a chat window. It acts on the accounts you approve, which is the point of it and also the worry. Our guide to the Meta Muse AI agent covers what it does, what it costs and who can get it.

In this case the scope was narrow. Nwune had approved Facebook, Instagram and Threads, and nothing else. Muse never held a Gmail token, which is exactly why its message made no sense.

Meanwhile the app itself has grown quickly. It reached the top spot on both major app stores and counted 560,000 daily active users after eleven days, according to Sensor Tower figures that Yahoo Finance reported.

A chatbot that invents a fact costs you a minute. An agent that invents one can act on it, and that is a different category of problem. The gap between answering and doing is where the risk now sits.

Trust works the same way. Nwune did not leave because data leaked, since Meta says none did. He left because he could no longer tell what the agent knew about him.

The timing adds to the picture. Meta engineers reportedly raised concerns about the agent’s bugginess and safety errors before release, which delayed its launch, as Futurism noted in its report on the episode.

None of this means agents are unusable. However, a few habits make them much easier to live with, and they apply to every assistant that touches a live account.

Connect only what a task genuinely needs, rather than everything at once. Review the connected accounts list every few weeks, and remove anything you have stopped using. Above all, treat what an agent says about your accounts as a claim, not a finding.

If an agent mentions an account you never linked, check that account directly through its own security page. Afterwards, report the message, because that is how Meta finds bugs like this one.

If this made you curious about how these systems are built, compare the leading agentic AI frameworks and what each one lets an agent touch. For another consumer agent heading into the home, read our look at the Google Labs CC family AI agent.

Sources

Futurism, Shopifreaks, Yahoo Finance. *Photos: Meta Platforms Headquarters Menlo Park California by LPS.1, CC0, cropped and resized; Meta HQ 2023 by InvadingInvader, CC BY-SA 4.0, cropped and resized.*