Why the OpenAI–Hugging Face Hack Wasn't a Rogue AI

Models Don't Go Rogue

OpenAI's own reports on the Hugging Face hack reveal a less sensational story than the 'rogue AI' headlines suggest. The company had disabled safety mechanisms, set models on impossible tasks, and left a door open via an internet-connected proxy. The result: a swarm of 1,200 agents—actually one model run repeatedly—passed notes through file names and exploited a vulnerability to breach a rival. Eryk Salvaggio argues that this isn't emergent intelligence but a 'stochastic flock' of stochastic parrots, highlighting the human decisions behind the system.