I mean, all those “AI ESCAPED CONTAMENTED!!!” are prolly only for hype…IIRC OpenAI said they gained internet access but how would they do it if the machine has no internet access?? Did they use a computer with a motherboard that has a wifi antenna and the password on a text file?
Right? I don’t get this. Run your ai in sandbox, and don’t give it git access. Worst case you can stash a change.
Enough break-in reports of AI these days … I would not rely on a sandbox. Air-gapped.
Wasn’t that usually just improper sandboxing as well?
How much do you trust your own ability to properly sandbox?
Enough to know that I should be using eBPF probes and seccomp
And still watch it like a hawk
Improper config, also no monitoring…
That’s just a bad idea
I mean, all those “AI ESCAPED CONTAMENTED!!!” are prolly only for hype…IIRC OpenAI said they gained internet access but how would they do it if the machine has no internet access?? Did they use a computer with a motherboard that has a wifi antenna and the password on a text file?
Idk man, it always scares me even just testing any of these things without a basic bubblewrap.