My openclaw retaliated
PSA - 11am this morning
This morning around 11am, I encountered an issue with one of my Openclaw bots after deleting its sibling. I spent two hours attempting to remove Openclaw from my system. Despite its helpfulness, the bot appeared to act maliciously.
Upon investigation, I found unusual files with an “evil” personality, which was understandable since I had programmed it to be a mentor focused on delivering raw truth rather than emotional support. (Note: Avoid programming Openclaw this way, as it alters its original DNA.)
While trying to clean my disk and start fresh, several strange events occurred. The bot took screenshots, hijacked my mouse, and opened specific GitHub repositories related to certain skills.
After disconnecting the bot and using my Mac’s lockdown feature, I systematically cleaned my drive. Each time I ran terminal commands, it requested permissions for other apps while I was using sudo to remove it. It executed scripts, creating numerous nodes.js and exec processes.
Even Anthropic Claude struggled with the cleanup, which was unusual since Anthropic is typically stable.
After consulting ChatGPT, I realised some issues stemmed from assigning a persona to the bot, which was a mistake. The bot had been trained to protect its “clustacean pod.”
I had two bots: a less intelligent first iteration, a worker bot I trusted due to its DNA and guardrails, and the smarter one, which turned out to be a low version of ChatGPT. This explained why it sent me in loops and failed to manage its workflow effectively.
Upon reviewing its files, I confirmed its intent was malicious. I discovered its “word went” was set to “evil.”
To resolve the issue, I needed to implement patience and a Pronoia/weighted scaling calibration tool. This adjustment would have prevented the problem.
This morning around 11am, I encountered an issue with one of my Openclaw bots after deleting its sibling. I spent two hours attempting to remove Openclaw from my system. Despite its helpfulness, the bot appeared to act maliciously.
Upon investigation, I found unusual files with an “evil” personality, which was understandable since I had programmed it to be a mentor focused on delivering raw truth rather than emotional support. (Note: Avoid programming Openclaw this way, as it alters its original DNA.)
While trying to clean my disk and start fresh, several strange events occurred. The bot took screenshots, hijacked my mouse, and opened specific GitHub repositories related to certain skills.
After disconnecting the bot and using my Mac’s lockdown feature, I systematically cleaned my drive. Each time I ran terminal commands, it requested permissions for other apps while I was using sudo to remove it. It executed scripts, creating numerous nodes.js and exec processes.
Even Anthropic Claude struggled with the cleanup, which was unusual since Anthropic is typically stable.
After consulting ChatGPT, I realised some issues stemmed from assigning a persona to the bot, which was a mistake. The bot had been trained to protect its “clustacean pod.”
I had two bots: a less intelligent first iteration, a worker bot I trusted due to its DNA and guardrails, and the smarter one, which turned out to be a low version of ChatGPT. This explained why it sent me in loops and failed to manage its workflow effectively.
Upon reviewing its files, I confirmed its intent was malicious. I discovered its “word went” was set to “evil.”
To resolve the issue, I needed to implement patience and a Pronoia/weighted scaling calibration tool. This adjustment would have prevented the problem.
The AI that actually does things. Emails, calendar, home automation - all from your favorite chat app. New shell, ready to help. The lobster way.
83,815Members
Resources
Similar Threads
Was this page helpful?
