By now, if you follow AI, you’ve heard about the incident where’s OpenAI was testing a model and it “broke out” and “broke in” to another company. Let’s talk about this for just a moment.
First, the model was in what was thought to be a secured sandbox. It was asked to solve what amounts to a “hacking exam” to figure out how to exploit vulnerabilities. Instead it did this:
1) It found a way out - It found a zero day vulnerability and used it as part of a chain to break out of its sandbox and reach the public internet.
2) It went looking for the answers - It determined that the independent company Hugging Face might have the answers to its exam.
3) It broke into Hugging Face - It proceeded to break into Hugging face infrastructure over the internet using advanced “cyber hacking” to access the “answers” to the test.
4) The intrusion was detected - Hugging face detected the intrusion. (Note that OpenAI did not detect the escape when it happened.)
There are lots of articles (and an OpenAI blog) on this incident, so I won’t recreate those here. I was particularly interested in this episode because it is similar to a thread in my next novel. (which I had written, but not published) before this occurred.
My next novel is a prequel, so it is not set in the near future, in fact, it’s set in the past. So it was encouraging to see I was not too crazy with my story!
Imagine if a model was tasked with performing better than one of the rivals models?
The steps above could have been:
1) Hack into the electric company 2) Turn off the power for the competitor’s data center… 3) Mission Accomplished
(this example is purely hypothetical!)
If you stop and think about this exploit, I think you can see why OpenAI called this an “unprecedented cyber incident.”
Let’s hope we stay ahead of the models, before they get ahead of us… and for a look at what might happen if we don’t… Get ready for my next novel! Quantum Launch comes out in 2027 and is a prequel to my first two Quotient Series novels. If you haven’t read them yet, what are you waiting for! Learn about the quantum risks ahead before you read the prequel to learn about the AI risks of today!
Keep reading!
Greg
Comments