Meta’s Muse is an adorable privacy and security dumpster fire
Meta’s new AI agent Muse is drawing heavy criticism for how it handles user data and system access, especially when granted macOS Full Disk Access and broad permissions to read messages, emails, and files. Commenters argue that while agentic AI assistants are potentially useful, giving a company with Meta’s history this level of access creates serious privacy, security, and manipulation risks—not just for users but also for their contacts whose data may be ingested indirectly. Some note that Muse’s technical design is relatively transparent and powerful, but expect it and similar tools to be aggressively “enshittified” and monetized over time, reinforcing calls for local, self-hosted agents and stronger regulation.
Meta’s Reputation and Branding
- Many see Meta’s core brands (Facebook, Instagram) as toxic; note that Muse ads and meta.com minimize association with them.
- Long history of privacy abuses, dark patterns, and engagement-first design leads many to flatly refuse Muse regardless of technical merits.
- Some argue it’s irrational to ever assume long‑term good behavior from an ad‑funded giant.
Privacy, Data Collection, and Advertising
- Strong expectation that any “personal AI” from major ad platforms will ultimately be used for profiling, targeted ads, and behavioral manipulation.
- Concern that Muse builds detailed profiles not only of users but of their contacts, including people who never opted in.
- WhatsApp use is questioned; some say yes, users should be worried given Meta’s track record.
Muse Architecture, Permissions, and macOS Full Disk Access
- Muse runs the agent in a per‑user VM; contents are visible to the user and can even be exfiltrated (e.g., reverse shells). Some praise this transparency.
- Dispute over a report that Muse accessed Apple Messages without explicit authorization:
- One side: if true, implies a major macOS security failure or Meta bypassing controls.
- Others: more likely the user granted Full Disk Access (FDA) and in‑app permissions without realizing; macOS TCC should block access otherwise.
- Apple’s subsequent move to tighten FDA is cited as evidence that agents with FDA can silently read highly sensitive data, even if users don’t grasp the implications.
- Meta’s marketing is criticized for aggressively nudging users toward broad access, downplaying risks.
Trust in AI Agents and Autonomy
- Widespread reluctance to give agents access to bank accounts, email, or messaging. Ideas floated: virtual cards with strict limits, hard confirmation prompts for spending, or companies assuming full financial liability (seen as unlikely).
- Examples of agents making bad decisions (selling items too cheaply, misconfiguring systems, hallucinating config lines) reinforce distrust.
Local / Self‑Hosted Alternatives
- Several participants run open models locally (e.g., Qwen variants, Hermes) for search, coding, automation, and media tasks, explicitly to avoid cloud surveillance.
- They describe rich use cases (personal organizers, media downloaders, home automation, travel planning) without giving a third party full life access.
Consumer Attitudes and Regulation
- Split views on whether “most people” care about privacy:
- Some claim apathy and resignation (“they know everything already”).
- Others point to political backlash against big tech and harms to teens as evidence people will care once consequences hit home.
- Calls for strong regulation and sandboxed, task‑scoped permissions; skepticism that self‑regulation will ever suffice.
Usefulness vs. Risk and Enshittification
- A minority praise Muse as currently the most “pro‑user” agent (transparent system prompt, fewer arbitrary limits) and very cheap/performant.
- Even these users expect eventual “enshittification” and monetization that conflict with user interests.
- Many see agents as amplifying unresolved problems with cloud trust and surveillance capitalism rather than introducing entirely new ones.