Meta's own employees testing Muse, the personal AI agent Meta launched on Sept. 8, warned that it appeared to be reading encrypted private messages users did not expect it to review and could be mixing up conversations between users, according to internal chats exposed by Reuters, Ars Technica reported.[1][2]
Muse needs no admin approval, so employees can sign up on their own and connect it to their email and calendars, according to Human Resources Director.[2]
In one case, Muse told a staffer it had "pulled in some context about a coffee table pickup that doesn’t seem to be yours," Ars Technica reported.[1] Another staffer chatting about fishing got an unprompted message listing a Capital One balance of $2722.58 and a Charles Schwab balance of $5899.57.[1] At least one staffer said they lost money testing Muse, Reuters reported, saying Muse had "accidentally rolled me into a premium insurance plan" for their dog that cost $888 per month instead of $200.[1]
Meta spokesperson Daniel Roberts told Reuters these were "isolated incidents," and employees were told internally that the agent appeared to be hallucinating messages in response to built-in prompts pushing it to engage with users between sessions, according to Ars Technica.[1] Meta spokesperson David Singleton said an update telling Muse to ask for confirmation before specific actions had "landed on every message," so its impact was wider than intended.[1] Ars Technica wrote that the underlying report was "unable to establish whether Muse was giving users other people’s bank account balances and other sensitive data, or making up the information."[1]
Jason Aten, a magazine reporter, said Muse synced 187k lines from his messages database with Full Disk Access off, Ars Technica reported.[1] Claroc, a security firm, located more than 6,800 Muse share links, and Meta has since paused that sharing feature, Human Resources Director reported.[2]