OpenAI says GPT-6 Astra is better aligned and more resistant to jailbreaks, but its safety testing also found reduced monitorability and some ability to evade oversight under adversarial conditions.