I got tired of all the AI agents that need access to my whole system so I built a fully sandboxed one
Posted by Poromenos@reddit | LocalLLaMA | View on Reddit | 8 comments
LionStrange493@reddit
Sandboxing makes a lot of sense for system access.
Curious though, have you noticed anything weird when the input itself tries to mess with the agent’s instructions?
Like not escaping the sandbox, just changing behavior through how things are phrased.
Poromenos@reddit (OP)
I haven't tested this because only the owner can talk to the bot, so if the owner wants to mess with it, that's acceptable. Subagents (to talk to other people) are separated and sandboxed themselves.
LionStrange493@reddit
yeah that’s fair, i’ve seen cases where even trusted input can still nudge behavior in weird ways though, especially once things start passing context around, not breaking anything, just… subtle shifts
Gregory-Wolf@reddit
Omg, is it called docker?
Poromenos@reddit (OP)
Hey, can you explain what made you decide to take time out of your day to make fun of the thing I made?
Gregory-Wolf@reddit
Sure. You were totally asking for it with "I got tired of..." and a product - these are being published every day here in gazillions.
You expected someone to take time out of their day with your post. It just went not how you planned. ;-) Unless you were trolling, then well-done!
But hey, at least you got my joke!
Poromenos@reddit (OP)
To qualify as a joke it has to be funny first.
McSendo@reddit
Like if you read it out loud, it does sound kinda ridiculous. Like your main product differentiator is to dockerize your agent.