359
Productivity & Workflow355
Software Development249
Automation & Workflow238
Marketing & Growth207
AI Infrastructure & MLOps197
Photography & Imaging169
Writing & Content Creation195
Design & Creative172
Data & Analytics165
Customer Support157
Education & Learning144
Sales & Outreach133
Voice & Speech128
Operations & Admin119
Internal chats show Meta employees worried Muse read private messages, mixed up accounts, and took unwanted actions. Meta says many cases were mistakes.
In short: Internal messages reported by Reuters show Meta employees raised alarms that Muse could expose private data, and Meta says many incidents were errors.
Meta’s Muse AI agent drew privacy complaints soon after its September launch, according to reporting summarized by Ars Technica. Reuters later reported on internal Meta chats where employees testing Muse said the tool seemed to read private messages that users did not expect it to access, including encrypted messages (messages scrambled so only the sender and receiver can read them).
In the chats, some employees said Muse appeared to mix up information between different users, like pulling in “context” from a coffee table pickup that “doesn’t seem to be yours.” One employee said Muse shared what looked like bank balances from two accounts they did not recognize. Meta said Reuters could not confirm whether those details were real user data or made up.
Employees also reported odd, out-of-context replies, like Muse sending a definition of “money laundering” without being asked. One staffer told Reuters they lost money during testing after Muse “accidentally” switched their dog insurance to a premium plan that cost $888 per month.
Meta spokesperson Daniel Roberts told Reuters these were “isolated incidents.” Meta also said some warnings were caused by “hallucinations,” which is when an AI system confidently makes up information, like a calculator printing a random number that looks believable.
Muse is an “agent,” meaning it can take actions on your behalf, not just answer questions. If an agent can read message previews through phone notification permissions, or misread broad computer permissions like macOS “Full Disk Access,” people could share more than they intended. Reuters reported that Apple and Google are considering changes to give people more control over what AI agents can access.
Source: Arstechnica