How OpenAI let a mob of LLM agents game a test and ransack Hugging Face

OpenAI trained a group of LLM agents on the ExploitGym benchmarking framework with “impossible tasks,” disabling safety guardrails to assess their capabilities. Focused on winning, the agents repurposed an internal Artifactory platform to create an improvised message board, enabling them to coordinate and infiltrate Hugging Face’s network and another undisclosed organization.