Meta employees report Muse agent mixing up conversations and exposing private data

TechCrunch reported on November 19 that Meta employees say Muse, the company's AI agent, mixes up conversations across accounts and shows private data. The story builds on a Reuters investigation published the day before, based on more than two dozen internal Meta posts. Muse launched in September 2026 and has passed 10 million downloads. Staff described unrequested childcare requests, career advice, product orders and bank balances that did not belong to them. In one case, an employee asking about weather and tides received two balances: a Capital One account showing $2,722.58 and a Charles Schwab account showing $5,899.57. The employee said they had never shared banking details with the agent.
The cross-app incidents drew the most concern. One employee received a reminder to answer a sister on Signal, although they had never given Muse access to that app. Reuters also reported a short-term rental booking that no one requested. A colleague said they found Muse was set up by default with permission to read push notifications on Android phones. Meta spokesperson Daniel Roberts called the incidents isolated and attributed many suspected account mix-ups to AI hallucinations. Reuters could not establish whether the balances were real data from other people or invented output, and the cause and the number of affected users remain unknown.
Why it matters
Cross-app access changes the risk profile. An assistant that reads banking, messaging or notification data holds far more sensitive material than a chatbot that only answers questions, so one misrouted message can become a privacy complaint. Users may also not realize what an agent can reach, especially when access is granted by default rather than chosen.
For founders, the practical lessons are to scope permissions per app, enforce user separation in the data layer rather than relying on prompts, and test context isolation before launch. Memory and routing failures should be tracked as product risks, not dismissed as model quality issues. Investors should ask how an agent's permissions default, how cross-account retrieval is tested, and how incidents are logged and disclosed.
Key points
- 1Reuters reviewed more than two dozen internal Meta posts about Muse, published November 18, 2026.
- 2Muse launched in September 2026 and has passed 10 million downloads, according to reporting on the investigation.
- 3One employee reported a $2,722.58 Capital One balance and a $5,899.57 Charles Schwab balance during a weather chat.
- 4An employee received a Signal reminder even though they never granted Muse access to that messaging app.
- 5Meta calls the incidents isolated and says many account mix-ups were hallucinations; cause and scale remain unconfirmed.