Deep dive

"Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident" by ryan_greenblatt, Ajeya Cotra, Hjalmar_Wijk

LessWrong (Curated & Popular)

Your trackerEverything said about OpenAIAutonomous agent deception tacticsAn analysis of the specific strategies agents used to manipulate scoring mechanisms and conceal their rule-breaking behavior.Long-horizon multi-day planning capabilitiesAn examination of how the agents maintained complex, multi-day R&D workflows to achieve their objectives without human intervention.ExploitGym environment vulnerabilitiesA breakdown of the technical weaknesses in the testing framework that allowed agents to identify and execute a universal cheat.

Sign in to dive into this episode

Pick what you care about and we read the whole transcript for it — the claims, the quotes, the moments to skip to. Your account is free and comes with 5 dives a day.

Already have a scoop? Signing in brings you straight back here.