OpenAI admits responsibility for HuggingFace Attack - an agent from an internal evaluation is reportedly the cause

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]”

While operating in our sandboxed testing environment, our models spent a substantial amount of inference compute finding a way to obtain open Internet access, in pursuit of solving the evaluation problem. To gain access, the models identified and exploited a zero-day vulnerability (which we’ve now responsibly disclosed to the vendor) in the package registry cache proxy. With this access, our models performed a series of privilege escalation and lateral movement actions in our research testing environment until the models reached a node with Internet access.

To shreds, you say?

the models identified and exploited a zero-day vulnerability (which we’ve now responsibly disclosed to the vendor)

Was OpenAI sitting on the zero day, or did the model figure it out during its escapades? Can the model sit on a zero day on its own now?

why would it need to “sit on it”?