I have a different interpretation of the OpenAI/Hugging Face incident than most of my friends in cybersecurity; I found it unsettling, and see it as an inflection point in the history of our field. Yesterday, the Wall Street Journal released an exclusive on Hacktron’s 3-hour automated path to initial access into OpenAI’s monorepo and internal employee tools; I did the technical due diligence on this piece for the WSJ over the past few days and was very unsettled by it; I see this as further evidence of a phase transition.
Anyway, I’ve been digesting all this, and wanted to write down my current set of non-theses and theses about where we stand in AI security: