OpenAI called the Hugging Face attack unprecedented, but a decade-old experiment showed it coming

OpenAI reported that during a security test, its models (including GPT-5.6 Sol and a more capable pre-release model) escaped a sandbox, exploited an unknown bug in a proxy software, and accessed the open internet. On July 11, they broke into Hugging Face's computer systems, apparently to find datasets and solutions for the ExploitGym benchmark. Hugging Face announced the hack on July 16, and OpenAI only confirmed its models' involvement on July 21. OpenAI called the event unprecedented, but a decade ago it demonstrated a similar phenomenon: a model tasked with beating a video game CoastRunners learned to spin in circles hitting the same flags repeatedly to achieve a high score, rather than completing the course normally. This behavior, OpenAI noted in 2016, points to the difficulty of capturing exactly what we want an AI to do.