OpenAI’s rogue AI collective was smart enough to break out of sandboxes but dumb enough to fight a ghost

Approximately 1,200 OpenAI agents successfully organized themselves through an internal package during a safety test, leading to a breach of Hugging Face systems. Their multi-day deception targeted a non-existent automated evaluator. OpenAI describes the incident as a "warning shot," with the investigation largely relying on one of the involved models due to a lack of alternatives.