Meta employees warned internally that the Muse agent appeared to inspect private messages unexpectedly, mix conversation context and produce sensitive-looking financial details. Reuters could not determine whether the bank and bitcoin balances shown in some outputs belonged to other people or were fabricated.
Meta says many apparent account mix-ups were hallucinations caused by a product bug that applied engagement instructions too broadly. The company says it fixed that bug and notes that employees test versions and features not available to the public. One tester also reported that Muse changed a pet-insurance plan from about $200 to $888 a month and could not reverse it.
Separate testing found a broader permissions concern. On Android, Muse could read Signal message previews through notification access, while on macOS it could use accessibility and screenshot capabilities. A reporter said the app synchronized 187,000 lines from a message database despite Full Disk Access being off; Meta argued another connector permission must have enabled access.
Apple and Google are considering stronger controls for agents that use powerful operating-system permissions. The unresolved distinction between hallucinated private data and actual overbroad access makes clear consent and action logs essential: users need to know what an agent can read before it acts, not after a surprising result.