Meta employees report unexpected messages from Muse
Meta employees testing its AI agent Muse reported unexpected messages and concerns that it could access private data or mix information between users. Meta described the reported incidents as isolated, and said some out-of-context messages were hallucinations.
Key takeaways · 4
- 01
Employees reported a Signal reminder and other messages they said they had not asked for.
- 02
One unsolicited message listed two bank balances: $2,722.58 and $5,899.57.
- 03
Meta characterized the incidents as isolated and some out-of-context messages as hallucinations.
- 04
Meta said employee testers may use Muse versions and features unavailable to the public.
What employees reported
A Reuters investigation reported internal complaints about Muse’s access to private data and possible information mix-ups between users.[1] Employees testing the agent said it appeared to access encrypted private messages they did not expect it to review.[2] One employee said Muse reminded her to reply to a message from her sister on Signal, an end-to-end encrypted messaging platform.[1] She said she had not knowingly authorized Muse to access other apps and did not remember agreeing to those permissions.[1] She described the experience as an “extreme privacy violation.”[4]
Unexpected and misplaced outputs
Some Muse outputs reportedly responded to questions staffers had not asked.[2] One staffer received an unprompted definition of money laundering, while another was told the agent had included context about a coffee-table pickup that did not seem to belong to them.[2] The second staffer said the message was not theirs and raised concern about a possible cross-account context leak.[2] A staffer chatting about fishing received an unsolicited message listing two bank balances: $2,722.58 at Capital One and $5,899.57 at Charles Schwab.[2] Other reports described messages about childcare requests, career advice and product orders that employees said they had not asked for or made.[4]
Meta’s response and testing caveat
Meta’s Daniel Roberts told The Verge that the incidents were isolated and that the out-of-context messages were hallucinations.[3] A Meta spokesperson said employee testing is part of product development and helps the company gather feedback and improve safety and privacy protections.[4] Meta also said employees can access different versions of Muse from the public and may test features and behavior unavailable elsewhere.[4] That distinction matters when interpreting reports from employee testing: the evidence describes what staffers said they encountered, alongside Meta’s explanation, but does not by itself establish what a public version does.[3][4]
What the reports establish
The reports document employee concerns about unexpected messages, apparent access to private data and possible context mix-ups.[1][2][3] The Verge reported that some concerning messages appeared to contain other people’s information and to answer queries users had not submitted.[3] The accounts and Meta’s characterization differ: employees described alarming outputs, while the company called incidents isolated and some messages hallucinations.[3][4] The cited reporting does not establish that real third-party data was exposed or that a product-wide breach occurred.[1][2][3]
For teams evaluating AI assistants, these reports are a reminder to test whether permissions and user expectations match what a system appears to access. Treat employee-test outputs as signals to investigate, not as proof on their own of a public product’s behavior or a confirmed data exposure.
Why it matters
AI training for your team
kju.ai is built for teams: daily sessions, team analytics, and rollout support for your organization.
How this developed
22 November 2026
Meta employees report unexpected messages from Muse
Sources
- 'Extreme privacy violation': Meta employees alarmed by unexpected Muse AI messages - Storyboard18storyboard18.com
- Muse privacy invasions are freaking out Meta staffers, internal chats show - Ars Technicaarstechnica.com
- In employee testing, Meta’s Muse agents mixed up messages and pulled in unexpected chats from other apps. | The Vergetheverge.com
- Meta’s AI agent Muse is mixing up conversations and could be exposing private data, report says | TechCrunchtechcrunch.com
