Red-team report found OpenAI utterly failed to monitor its agents during Hugging Face attack
OpenAI's monitor only looked in when an agent handed back control, so after a compaction an agent could go unwatched "indefinitely" by the simple method of never taking a break.