Meta's Muse, an AI agent only a few months old, can read the text of users' messages through permissions granted at setup, without explicit access to the messaging app, according to a Reuters investigation reported by ZDNET.[1] Meta employees testing Muse said the agent took actions they had not asked for and messaged them about topics including bank account balances they didn't recognize, ZDNET reported, citing a Reuters investigation.[1]

The usual advice for running an agent is to restrict which apps it can touch, but ZDNET argued that this kind of privacy hygiene is not enough if Muse can get in anyway.[1] ZDNET also noted that Muse reportedly hit 10 million downloads in just 40 days, and wrote that even if you don't use Muse yourself, a closer look at Meta's privacy policy "reveals how it can still access your information."[1]

Reuters and the AI Now Institute ran their tests on an Android phone and a Mac, according to ZDNET.[1] Because Muse is set up with automatic access to Android push notifications, it did not need explicit app access to read messages, while the same tests did not work on an iPhone.[1] One employee was notified about a Signal message, though she said she had never given the agent access to the encrypted app, and another had Muse book a non-refundable Airbnb he had not asked for.[1]

ZDNET reported earlier incidents: in September a user said Muse shared his address with a Facebook Marketplace buyer before checking with him, and in its first month a cybersecurity professional found a zero-day bug that made the agent hijackable.[1] ZDNET contributor David Gewirtz, who tested the agent, called it the worst AI agent for privacy.[1]

Meta told Reuters the incidents were either isolated or the result of hallucinations, though Reuters could not confirm that, ZDNET reported.[1] ZDNET suggested the findings could help Google's agentic offering, Tasks, formerly Spark, which Google made free with usage limits.[1]