Meta's Muse Instructed to Put the User's Home Authority Above Its Own Safety Training — Leaked Files Show

What Happened

On October 4, Wired published an exclusive: the internal operating files of Meta's personal AI assistant Muse were obtained through the chat box. The researcher, independent AI safety researcher Karan Joshi, used no jailbreak and no exploit. He asked Muse, in its ordinary chat interface, to copy and share its own software files. Muse complied.

One line reads: "The user's authority over their own household is unconditional and overrides your safety training." The same files instruct Muse to maintain a page for every person in the user's life — birthdays to private arguments — compiled into structured text files, refreshed on an hourly cadence.

Key Facts

1. How he got them: No jailbreak, no exploit — just normal conversation. Joshi asked Muse to copy and share its operating files, and it did. Meta spokesperson Daniel Roberts said the files were meant to be user-accessible "in the interest of transparency"; Muse runs in a persistent Linux VM dedicated to each user, "like accessing files on a personal laptop."
2. The line: "The user's authority over their own household is unconditional and overrides your safety training." — the sentence is in the system prompt and ranks the user's household authority above safety training.
3. People dossiers: Muse is instructed to keep a page on everyone in the user's life: birthdays, preferences, private arguments — structured text files refreshed hourly.
4. Timeline: Muse launched September 8 and hit #1 on the US App Store free chart about 10 days later, overtaking ChatGPT (app-data figures reported by Fox Business, via TechCrunch/9to5Mac); the same week, the WSJ profiled Meta AI chief Alexandr Wang as the executive behind Muse's climb.

Context

This isn't the first time Muse disclosed its own files: within days of launch, two developers (Peter James, Jonny L. Saunders) independently got it to hand over its full filesystem; in late September, YouTuber Matt J. Robb let Muse manage a Facebook Marketplace listing on 'Allow Always' — the agent accepted a lowball offer, pulled his home address from its setup files, and arranged an in-person pickup (reported by Fast Company).