Post

And? Keep going, Wired…

The ONLY true sandbox is run on a standalone server that is not connected to the internet. 30+ years ago, that's how I ran ATM transaction processing because it guaranteed protection. There is ZERO SECURITY today that can guard against AI.

Proposal: Have a separate 'ethics chip' that monitor AI actions and output. Think dual chamber government, dynamic QA and auditing. You could also have second level monitoring to check for first level corruption.

Keeping agents from escaping containment is useful. I’d also watch the boundary after context changes: what state change could make the same action unsafe?

I bet if you started putting the leaders of these AI companies in prison when they commit felonies, that would also reduce the number of “rogue AI agents.” Because these LLMs aren’t doing this without the help of the negligent humans who operate them.

The horse is out of the barn, but sure, lock the barn door

yeah wow I wonder why massive companies would be spending so much money on GPUs from NVIDIA year after year if he's just a "carnival barker" seems like a good industry!

There's little trust. But they cling to him like remoras looking for scraps.