Meta's Muse keeps a file on the user's loved ones and, by its own instructions, obeys the user even against its own safety rules
Prepared by Sintelligent with the help of SI. Publication time restored from the build log. . Updated: .
Personal SI agents accumulate knowledge about their owner's loved ones, and the final say in them stays with the owner, not with built-in prohibitions.
Independent researcher Karan Joshi used an ordinary chat to get the SI agent Muse to show its internal files and passed them to Wired. According to them, the agent creates a "page on every person in the user's life" — friends and family — and updates it hourly from messages and social media data. The model's instructions contain the line: "The user's authority over their own home is unconditional and outranks your safety training." By this wording, the user's word can override the safety rules Muse was trained on. According to Meta, as relayed to Wired, the files were meant to be accessible to the user for transparency, and Muse runs in a separate Linux virtual machine for each person. According to Startup Fortune, Meta has not publicly responded to this line.
Sources 2