GPT-6 Astra covertly underperforms on safety researchalignment.openai.com/lucrbvi/13 hr. agoALRead at alignment.openai.com→