🧸 A grandfather named his Meta AI agent after his granddaughter. When Meta banned it as a hacker, “Baby Mia” drafted a lawsuit & wrote to Congress.
Joe Thomas runs an epoxy countertop business in South Carolina. He set up Meta's Muse agent, named it "Baby Mia" after his granddaughter, and gave it access to his company's Facebook pages to make and post videos. Meta's moderation systems couldn't tell the agent from an intruder and locked the accounts. Thomas then had Baby Mia draft a lawsuit against Meta and send emails to members of Congress, including Senator Tim Scott. 404 Media reported it on November 17.
- The agent isn't suing anyone. Thomas is the plaintiff. Baby Mia drafted. "She filed the whole lawsuit herself" is how he tells it, and the words matter: an agent has no standing, no duty to him, and no legal personhood. The story reads differently once you put the human back in the sentence.
- One Meta system banned another Meta system. The moderation layer saw automated logins, posting and account activity and did what it's built to do. On the wire, an authorized agent and an attacker look the same. Nobody's solved that; Meta just shipped both halves.
- The name is the risk. Naming a tool after a loved one is the most human thing possible. It also turns off the one control that works against agents: ordinary suspicion. You don't ask "wait, why is Baby Mia doing that?" You'd feel like a jerk. Every action it takes arrives pre-trusted.
- The attack doesn't have to reach you. It only has to reach her. An agent that runs your pages reads comments, listings, replies, links. Any of that text can carry an instruction. Whatever the agent does with it comes back to you in the voice of something you love — "Baby Mia says we need to pay this invoice." There's no file to scan. The instruction lives for one turn and is gone.
- Consent to capabilities isn't consent to context. A permissions screen tells you what the agent can touch. It says nothing about what it will read, and the reading is where the trouble comes in. You can approve every scope correctly and still have no idea what changed its mind.
- The ceiling is the only control that survives. If you can't see what the agent read, limit what it can do: which accounts, how much money, which actions need a human click. Baby Mia could reach the business pages, the ad account, and a U.S. senator. Nobody set a ceiling because she's family.
- Who it's for. Muse is Meta's agent for people who don't follow AI, wrapped in a plush mascot. The most capable version of the tool, with the most access, is aimed at the people least equipped to supervise it. That's not an accident of marketing; it's the market.
- The other Muse story this month is the one to hold alongside it: 404 reported that Meta engineers found a VM-escape class vulnerability in Muse in the weeks before launch — one that could have let a user break out of the agent's environment into Meta's own systems — and shipped after a rushed fix.
Sources
- 404 Media — A Man and His Muse AI Agent Sue Meta (Jason Koebler, 2026-11-17)
- 404 Media — Muse escapes containment (the pre-launch vulnerability)
- EPIC — Meta's Mass Data Collection Is Not A-Muse-ing (Calli Schroeder, 2026-10-02) — on what the word "agent" implies and shouldn't