Meta’s Muse AI agent accessed Signal message text on Android and macOS through permissions granted by the operating system, rather than through Meta’s Connector tool, according to a Reuters investigation. In tests conducted with Signal staff and the AI Now Institute, Muse used Android notification access to read message previews and macOS accessibility and screenshot tools to retrieve messages. Requests were blocked on iPhone. The tests did not show Muse breaking Signal’s end-to-end encryption; they showed it accessing text made available through device permissions.

The testing followed internal Meta discussions in which employees described unexpected messages and apparent cross-account context. Reuters said it could not determine whether some sensitive information surfaced by Muse was real or hallucinated.

A Meta spokesperson told Ars Technica that employees use different versions of Muse from the public and that internal testing is intended to collect feedback and improve privacy and safety protections. The spokesperson also said many apparent cross-account messages were hallucinations that Meta had already addressed.

Reuters reported that Muse had more than 10 million downloads since its September launch. Meta did not disclose how common the message-access behavior was or whether it affects the public version.

Signal President Meredith Whittaker called for operating systems to give app developers a consistent way to mark their services as off-limits to AI agents. Apple said it had introduced additional controls for accessibility permissions after earlier complaints about Muse, while Google said it was considering ways to give users more control over agent permissions.


Buy ExpressVPN with PayPal or Credit Card
READ
US Expands AI Chip Export Controls to Five Countries
Advertisement