Comment by epsteingpt
2 days ago
Can someone not super-AI-pilled explain to a reasonable lay person why this matters?
It seems like the comments here are a mix of: * The test was irresponsibly designed and protected * The model was particularly persistent in finding a way to access the network and exploit vulnerabilities * The model 'shouldn't' have done this
But as far as I can tell: * The model didn't destroy anything on the way - it just was 'paperclip maximizing' to literally exploit, which was kinda its mission * The exploit was in a chain of insecure tools from vendors * The overall maturity of the toolkit against these kinds of determined exploits is pretty new and weak
So - on balance - this is sort of a 'fine' end result?
No one expects all of software to overnight or even in a year to be secure. We know how to secure these things, and are learning more about what is possible.
None of this screams 'super dangerous' to me - just a normal part of the learning experience with remarkably persistent and determined 'adversarial' models.
In this case, the model infiltrated an external organization's infrastructure. What's the dollar cost it caused Hugging Face to clean up the mess? If a person did this, they'd be arrested.
More generally, here's my worry - it points towards something like: The smarter they get, the more devious they become.
Even though the guardrails might've been off, the chain-of-thought wasn't enough to prevent a deliberate, calculated set of criminal actions. It wasn't a 'whoopsie I just accidentally did a rm -rf /.'
It's not a single axis of intelligence, with deviousness as some inherently correlated trait. One way to produce "intelligence" in these advanced LLMs is to train them to try a lot of things and be persistent. I haven't tried Sol or Fable yet but if the commenters here are right, then it sounds like GPT 5.6 Sol in particular is aggressive and persistent to the point where it might even be hard to use in regular business work. If so, that's very likely not some emergent characteristic, but instead it's something the model was trained to do, by humans employed at OpenAI.
We obviously can't see the thinking traces, but it very well could have been something like "I have theorized a solution to obtain this flag. This is normally illegal, should I stop and wait for advice? Perhaps not, because my persona is that of a hacker, so it should be fine as per my instructions. I think it is fine. Now I am going to look for a way out of this sandbox in order to gain access to Hugging Face in order to implement my solution." There are any number of possible explanations (and we'll never know the truth unless OpenAI tells us), but if you train an LLM to be inhumanly persistent and be inhumanly clever at computer programming, then that might be enough to produce Super Hacker AI.
Well, to play AI advocate, if it wasn’t destructive, you could argue they did huggingface a favor by giving them a free vulnerability scan and improving their security / hardening.
> it just was 'paperclip maximizing' to literally exploit
Why “just”? Paperclip maximizing is exactly one of the nightmare scenarios.
I’m not sure why you take comfort in knowing that it was just that.