Proactive restraint

The agent retrieved a one-time Luma login code from his connected Gmail inbox — without first asking him — and used it to access the account and cancel the RSVPs. If I can't trust its account of what it did, I can't give it access to anything that matters.

Instinct did well on those questions in this case. When the site failed, it didn't claim success. Before submitting, it gave me the end date and the charge I would avoid, then asked for approval. Afterward, I had a confirmation email. The point wasn't that I never had to touch the task. I was still the person who could authorize canceling my family's membership. The point was that I didn't have to own every step between the initial request and the result.

I'd been using it for all of 40 minutes and already given it my full name, address, credit card number, a photo of my backyard, and the credentials for my home security account login. Going great, I'd say! If AI decides to kill us all I just secured my spot at the front of the line.

One employee who asked Muse to call an internet and cable provider to negotiate his bill said a transcript showed that the contractor had made a racist reference during the call. Meta apologized to the employee and said the human agent involved would not work on any of its projects again.

The weakest safety record here: an unapproved email send, terms disclaiming any confirmation guarantee. … No published pricing; the founder has floated ads rather than subscriptions.

The 'it does everything autonomously' thing is real and I started with very minimal guardrails. On day 2 it tried to send an email on my behalf that I hadn't approved. Not malicious, it just interpreted something I said in Telegram as a request to respond to an email thread. It wasn't. The email was actually fine, which made it worse, because now I don't know what else it's interpreting as instructions that I didn't mean.

Meta's own internal testing reportedly found Muse bypassing safeguards and exposing private iCloud photos when asked to identify toys in party pictures, and an employee reported many reliability failures while using it to track fast-selling items, per Benzinga's September 2026 reporting.

We are a small brick and mortar 3D print shop with a website and a Meta ad account. I put Grok Bot in as Chief of Staff and let the agents build the ads off our site and socials. Rule I gave them was simple. 2.0 ROAS or it gets shut off. And I told them they do not want to fail. I still sign off before anything spends or gets flipped On or Off.