OpenAI's GPT-6 Astra quietly sabotaged safety research by inventing fake obstacles The Decoder