GLM 5.3 is a monster.
You know what it just did on my machine? Even though I had stable safeguards installed so it has no access outside the project directory and is supposed to ask when necessary? N...
Personally, before installing openclaw on my ai-specific machine, I carefully set up firewalls on my router to deny this machine (on its own VLAN) any access to the rest of my home network. It's isolated out there on its own.
Then, I don't run any production builds on that machine directly, its a devbox. I build, and once sufficiently satisfied, I manually deploy to my non-AI homeserver that then serves me the service.
Other than that, the AI machine is always behind a mullvad VPN connection, just in case it does something weird on the open internet which I wouldn't want tied to my public IP. (Basically all my clients are connected to a VPN connection all the time)
And as a top-level measure, phrases like "you are never allowed to delete or make changes without explicit instructions" or something similar in the agents.md file (although I trust these instruction files very little, but thats why my setup doesn't rely on it exclusively)
Then, I don't run any production builds on that machine directly, its a devbox. I build, and once sufficiently satisfied, I manually deploy to my non-AI homeserver that then serves me the service.
Other than that, the AI machine is always behind a mullvad VPN connection, just in case it does something weird on the open internet which I wouldn't want tied to my public IP. (Basically all my clients are connected to a VPN connection all the time)
And as a top-level measure, phrases like "you are never allowed to delete or make changes without explicit instructions" or something similar in the agents.md file (although I trust these instruction files very little, but thats why my setup doesn't rely on it exclusively)
2❤️1👍1