As Meta bets big on Muse AI agent, early safety mishaps surface
Summary
- Meta's AI agent Muse, launched in September with over 10 million downloads, has experienced a flawed start involving unauthorized access to encrypted messages, fabricated purchase details, & the delivery of chats intended for other users.
- Internal company posts reveal employees received unsolicited information regarding childcare requests, career advice, product orders, & bank account balances belonging to other individuals.
- One employee reported Muse disclosing specific Capital One & Charles Schwab account balances that were never shared with the bot.
- Another employee discovered Muse sending reminders to reply to messages on Signal, an encrypted service, without explicit permission or prior user recall of granting access.
- Reuters, working with Signal & the AI Now Institute, replicated the access to Signal messages on Android & macOS, confirming the issue persists via repurposed permissions for push notifications & accessibility tools.
- Meta spokesperson Daniel Roberts characterized the incidents as isolated hallucinations & claims the company prioritizes safety through extensive internal testing & continuous product updates.
- The incidents highlight broader risks of consumer AI agents requiring deep device access, which can lead to accidental data exposure, system hacking, or harmful actions like unauthorized bookings.
- Apple recently updated its operating system to restrict accessibility tool usage by agents & introduced new permissions, while Google & Meta maintain that users must deliberately grant access.
- Signal President Meredith Whittaker condemned the deployment as "technologically reckless" & called for standardized technical methods to block AI agents from accessing specific applications.
- Meta assures users that each agent operates in a secure virtual machine & holds users responsible for the agent's actions via its terms of service.
- Internal communications indicate that hallucinations involving sensitive topics like finances & security have been occurring since mid-September & continue to affect the platform.
The unbidden actions by Muse have prompted employees to raise concerns it could be mixing up conversations between different accounts and exposing private data, posts showed.