SITUATION EXPLAINED: OpenAI explains at Black Hat how its agents ended up hacking Hugging Face. • The models had no internet access, only a package manager called Artifactory, which did • Agents stuck on impossible tasks uploaded notes to Artifactory hoping others would read them, turning it into a shared message board • One agent found a server-side request forgery on Artifactory and shared it, so other agents could reuse the exploit • What followed is what OpenAI calls a Cambrian explosion…

Greg Brockman@gdbAug 6

Black Hat talk from the team, with a detailed timeline of and takeaways from the OpenAI-Hugging Face Incident: https://www.youtube.com/watch?v=87DyyMV0kCY