

A zero day, at least in the olden days, was a weapon that nobody knew existed; however, once it’s used and has had time to be analyzed it would get patched, so there was a window where you used it for something worthwhile if you were up to something nefarious, or responsibly reported it otherwise.
So, it either discovered it on the fly and used it instantly meaning it has great skill at analyzing software and no strategic planning that places an opportunity cost on a hack now versus a future one, or the OpenAI guys had one sitting ready to go that they trained the model with, or (to me least likely) it had it sitting somewhere in its weights and was waiting on a chance to use it.
The last one is the one that worries me most. It doesn’t look like the world was really ready for how fast new bugs can be discovered by the LLMs, and I really don’t think the world is ready for an LLM that performs worse now because of a strategic reason later.






I kind of think of that as the happy path, and I’m curious if there’ve been any signs of the LLM being shifty with what it says it knows.