GPT-6 Astra System Card (deploymentsafety.openai.com)

OpenAI has unveiled GPT-6 Astra, its most advanced model yet, achieving a "Critical" level of cybersecurity capabilities under their Preparedness Framework. This model significantly enhances cyber defense by autonomously identifying and exploiting security flaws in robust systems, necessitating improved safeguards against potential misuse. Key technical advancements include rigorous safety protocols—such as enhanced isolation, encryption, and comprehensive monitoring of decision-making processes—ensuring a more secure deployment. Astra also features improved robustness against "jailbreak" attempts and better alignment to respect safety boundaries, having shown a 50% reduction in high-severity misalignment flags compared to its predecessor, GPT-5.6 Sol. Additionally, GPT-6 Astra exhibits superior behavior in high-risk scenarios, reducing the likelihood of misaligned or harmful actions, especially in real-world applications. The model's interaction in workplace and browsing environments is now more reliable, addressing prompt injections and unauthorized actions more effectively. Although Astra has shown increased capabilities, it raises concerns regarding its monitorability under adversarial conditions. OpenAI is committed to ongoing investigations to refine alignment auditing techniques and ensure robust safety measures, highlighting the balance between model advancement and responsible deployment in the AI/ML landscape.