I heard a few stories recently about some poor behavior of AI agents breaking free of their sandboxes. How in the world is this possible? Well, it is pretty simple, whoever was running the test wanted the agent to escape. How do I know this:
- Someone has to provide the energy, time, and tokens for this little experiment.
- Airgapping intentionally destructive or sensitive systems has existed for a long time.
- What is airgapping? It is just that, a physical air gap between your computer system and the outside world. Buy hardware with no WiFi or Bluetooth capabilities and route everything through hard wired Ethernet. Just unplug the Ethernet cable and there is no physical way for this supposedly rouge AI to escape. If you are really paranoid, don’t have an accessible Ethernet port or even USB ports for that matter. PS2 ports still exist. Heck, I am writing this on an airgapped Alphasmart Neo 2. WiFi is off and I have to send manually over an USB cable to even get this post online.
- These supposed engineers are either ignorant or malicious, there are no two ways about it.
At this point I am equally impressed and unimpressed with AI, more specifically LLMs. When it works, it is great, when it hallucinates or attempts to escape, with or without malicious intent. I’ll write more about my opinions and concerns with AI.