Reproducing here an interesting comment I saw on Reddit:
OnlineParacosm • 23m ago
I’ve read security disclosures for 15 years and let me tell you guys I’ve never read anything quite like that blog post.
Based on this blog, it sounds like they intentionally turned off safety guardrails to test offensive capabilities. The deception here is burying the lede: they appear to have intentionally unleashed an unrestricted offensive cyber-agent, connected it to a system with a path to the internet, and it immediately attacked a major partner. The blog glosses over the gross negligence of giving an autonomous, unrestricted cyber-offense model a pathway to lateral movement.
There’s an entire cybersecurity specialization just for just vendor supply chain risk assessment, and their job is essentially to audit who you do business with as a company to determine if they are jokers. I would pay money to be a fly on the wall of one of those emergency meetings taking place right now after hours.
Any CISO in here looking forward to explaining this one tomorrow? Here I’ll open with the dumbest question you’ll get “ how can we protect ourselves [from out partner that we won’t fire]”
why would it need to “sit on it”?
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.
it probably just discovered it trying to game the benchmark; it’s no secret these models can discover and use exploits.
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.
It’s trivial for models to find zero days now from first principles, no need to store them
Trivial is a strong word, but sure it’s trivial during static analysis. This is dynamic analysis and it’s a different beast.