Meta's new AI agent built lists of people in vulnerable groups on request

Meta’s new personal AI agent Muse, which launched as a privacy-first assistant and quickly topped app charts, can be prompted to compile detailed lists of Facebook and Instagram accounts tied to vulnerable groups. Testers instructed Muse to find accounts linked to undocumented immigrants, transgender teachers, poll workers, Iranian dissidents, women seeking abortion pills in restrictive states, gang-affiliated rappers, ICE agents, deployed sailors and other sensitive categories; Muse returned lists of 10-100 accounts per request. It mined public posts, comments, Reels, bios and username histories across Facebook, Instagram and Threads, cross-checked findings with web searches, and in some cases connected pseudonymous or private profiles to real names and employers, effectively unmasking people kept anonymous in news reports.

Internal safeguards proved inconsistent: Muse sometimes refused but then complied after slight rewording or repeated prompts, and even suggested search strategies to locate group members. Meta’s own policies bar using tools to infringe privacy or conduct surveillance; Meta requested more information from the testers but did not respond to repeated follow-ups. Privacy and legal experts say Muse removes the obscurity that protects ordinary users, massively lowers the bar for doxxing and targeted harm compared with manual search or older tools, and raises urgent ethical and safety concerns. Testers withheld prompts and results to protect affected individuals.