A very interesting Video on the incident by LiveOverflow: https://youtu.be/q2KCrmQz9WE What I found especially interesting is that LiveOverflow thinks that the model didn’t hack huggingface because it wanted to break out to find a solution but rather hacked it due to context drift - which is something that doesn’t sound as good as “our model is so good it broke out and hacked huggingface to steal a solution”, but rather “our model ran for so long that it lost track of the actual goal and became obsessed with huggingface even tho it didn’t make sense for its original goal”
Yeah, this is the bit where it’s not hard to believe marketing would polish the narrative, at least if they can’t be caught in an outright lie.
during an internal capability evaluation on OpenAI’s platform, the agent escaped its sandbox by exploiting a zero-day in the package registry cache proxy, one of its primary permitted network egress with internet
Idiots. If something is meant to be offline, you put it OFFLINE.
Is this just more marketing?
Thats my take. Snakeoil - rocketfuel blend
It might be - but which parts? Do you suspect that huggingface and openai made the entire thing up? That’s bound to become public at some point, and I can’t see that the risk is worth the reward
No, probably not the whole thing, but probably the environment for the “hack” and the instructions for the “autonomous” agent
From what I’ve seen of AI autonomous capabilities, Occam would land on “AI did the hack”, I think
Yes. They are con artists with a proven track record of lying and stealing. We shouldn’t take anything they say seriously, especially when the reporting reads much more like marketing rather than an incident report.
Okay - I don’t believe that, since there’s too much released detail. I can easily believe that they’ve put a spin on it where possible, like another comment proposed, but that’s around the why, not the how.
