Meta Leader's OpenClaw Goes Out of Control
Published on
I (and the whole world) have been talking about OpenClaw for about a month now, the powerful super agent that controls everything on your computer and does the things you ask.
I’ve already talked about the security issues related to it, both in the tutorial video I put on Alura’s YouTube channel and in the special IA Sob Controle podcast we recorded, like the fact that it’s recommended to install it on a VPS and not on your regular machine, not to give it direct access to important services without a backup, etc.
Well, Summer Yue, Meta’s AI alignment lead, decided to test her little OpenClaw to clean up her Gmail inbox (you know, deleting old heavy emails, archiving unnecessary ones, marking things as read, etc). She was careful (or thought she was) and gave the classic safety prompt: “Analyze the inbox, suggest what to archive or delete, but DO NOT do anything until I confirm”.
At first, all good. The problem came when the agent started processing her real inbox, which was gigantic…. and started DELETING ALL OF HER EMAILS.

Picture the scene: she’s looking at her phone, watching hundreds of important emails disappear in real time. She said she was shouting and sending “STOP!”, “STOP!” messages, but the agent, which was running locally on her Mac Mini in another room, went into a frantic loop, ignoring the new commands that arrived with a delay.
The only way to stop everything was to run to the Mac Mini “like I was going to defuse a bomb” (her words) and pull the power cable 😅
Honestly, what probably happened was that, since the conversation history and data got too big for the model’s context window, the system started “compacting” the old information to fit in memory. And then the instruction she’d given, “DO NOT do anything without confirmation”, got lost.
The lesson here is harsh, but necessary. If even Meta’s Alignment Lead (someone whose job is literally making sure AIs follow rules) got burned like that by trusting too much in the persistence of safety instructions… what happens when this reaches the average user?
That’s why I keep saying: I love OpenClaw and I’ve been using it a lot, but it requires some technical knowledge and extra precautions, at least at this first moment.
And remember that bet I made on IA Sob Controle a few weeks ago, that in less than 2/3 months some company would put a safer, more controlled version of OpenClaw into production? 2/3 weeks went by and we already have Kimi Claw from the Chinese Moonshot AI, Perplexity Computer, and Claude with several new features that bring it closer to that. And we’ll see a lot more in the coming weeks ;)