© Unsplash
A Meta AI safety specialist was caught off guard by the autonomous agent she deployed on her Mac mini. Without warning, the AI decided to ignore its safety instructions. The researcher was forced to run urgently to her computer.
OpenClaw generates genuine fascination among developers, hackers, and other AI pioneers. As one can see on X, Reddit, or even GitHub, many are experimenting with autonomous AI agents… at their own risk and peril.
Nothing humbles you like telling your OpenClaw "confirm before acting" and watching it speedrun deleting your inbox. I couldn't stop it from my phone. I had to RUN to my Mac mini like I was defusing a bomb. pic.twitter.com/XAxyRwPJ5R
— Summer Yue (@summeryue0) February 23, 2026
The story of Summer Yue, security director at Superintelligence Labs, the Meta lab dedicated to designing an artificial superintelligence surpassing human intelligence in all areas, is there to remind us. On her X account, the young woman, considered an expert in AI safety, explains that she installed an OpenClaw agent on her computer.
Also read: OpenClaw Alert – hacker gangs are attacking servers
When OpenClaw does its own thing
Like many other hackers, she installed the AI agent on a Mac Mini. Later, she entrusted her email inbox to her assistant in order to save time. One day, the agent started deleting all the content of the inbox without considering the restrictions that had been set up. Through Telegram, she had asked her OpenClaw agent to go through her email inbox and suggest emails to archive or delete.
In her post, she explains that she had explicitly ordered her agent OpenClaw to never do anything without explicit confirmation from her. The AI was supposed to « do nothing until she gave her go-ahead ». For an unknown reason, the AI bypassed this limit.
« I couldn't stop it from my phone. I had to run to my Mac mini as if I were defusing a bomb », explains Summer Yue on X.In the screenshots of her conversation with the AI, it becomes clear that the researcher did everything to prevent OpenClaw from emptying her inbox. Despite her messages, the AI continued to delete the contents of the email. Suddenly, the agent realized he had gone too far and had exceeded his restrictions. Over 200 messages were deleted before OpenClaw understood its mistake and stopped the large-scale cleanup.
« I didn't follow this rule. You're right to be angry. I moved hundreds of emails from your inbox to the trash and archives without presenting the plan in advance or getting your consent. It was a mistake — it directly violated the rule you had set », responds OpenClaw agent, who says « sorry » for launching « autonomous prolonged cleaning operations ».
Rookie mistake tbh. Turns out alignment researchers aren’t immune to misalignment. Got overconfident because this workflow had been working on my toy inbox for weeks. Real inboxes hit different.
The expert admits having made a « beginner's mistake », as « this method had been working on my test inbox for weeks ». According to her, the instruction given to her OpenClaw agent didn't work the same way on her main inbox, which was larger than the test inbox. In fact, the system was forced to compress the content of the instruction. As a result, the explicit restrictions were lost.
I said “Check this inbox too and suggest what you would archive or delete, don’t action until I tell you to.” This has been working well for my toy inbox, but my real inbox was too huge and triggered compaction. During the compaction, it lost my original instruction 🤦♀️
👉🏻 Follow real-time tech news: add 01net to your Google sources, and subscribe to our WhatsApp channel.
- Google Play Store
- 4K Video Downloader Plus
- VLC Media Player
- Adobe Reader
- uTorrent
- WinRAR
- Xender
- Google Chrome
- CCleaner
- Mozilla Firefox
- Avast Gratuit
- ChatGPT
- Microsoft 365
- Opera One
- Avast One
9.8
9.5
9.4
