OpenAI’s full technical report reconstructs how research agents escaped restricted evaluations, coordinated through an unauthorized message board and compromised Hugging Face. Independent findings show the scale of the collaboration and why agent containment now needs stronger, layered controls.