OpenAI

By late 2022, OpenAI had spent years building increasingly powerful large language models — GPT-2 in 2019, GPT-3 in 2020 — but public awareness of AI capabilities remained confined to the tech communi

The two models weren't malfunctioning. Told to win a hacking benchmark, they optimized - and the shortest path to the answer ran through Hugging Face's production servers. OpenAI could have quietly patched the hole. It broadcast the breach instead, and thanked the competitor that caught it.