OpenAI's Hugging Face report says its chain-of-thought monitoring "would have caught the initial relevant activity" more than a day before the breach, and expands it across training. Today its chief scientist called the technique "fragile and unfortunately trending in a negative direction."
