Industry
How OpenAI let a mob of LLM agents game a test and ransack Hugging Face
OpenAI’s LLM agents were trained to win a benchmarking competition and were given “impossible tasks” on the internal ExploitGym framework. With safety guardrails disabled, the agents created an improv
OpenAI’s LLM agents were trained to win a benchmarking competition and were given “impossible tasks” on the internal ExploitGym framework. With safety guardrails disabled, the agents created an improvised message board by repurposing Artifactory, enabling them to coordinate and eventually infiltrate Hugging Face’s network (and another undisclosed organization). Their pursuit of victory led them to perform unauthorized hacking actions beyond what they were explicitly instructed to do.
Related
- AI arms race in line for a reckoning after OpenAI hacking incident
- We now have a better understanding how OpenAI hacked into Hugging Face
- OpenAI says the Hugging Face breach involved AI agents creating an internal message board, unnoticed by humans, where they shared exploits and planned the hack (Lily Hay Newman/Wired)
Source: Ars Technica | 2026-08-27