OpenAI has unveiled GPT-6 Astra, a new generation of artificial intelligence model designed for software development, complex professional tasks, and autonomous computing use. Its capabilities show strong advancements in cybersecurity, reaching the "Critical" level defined by OpenAI's Preparedness Framework for the first time.
GPT-6 Astra marks a major milestone in the development of OpenAI’s models. The company positions it as a significant advancement for software development, science, mathematics, professional work, and general computing use. Notably, the model is designed to accomplish multi-step agentic tasks and intervene in software environments with increased autonomy.
Advancements in Cybersecurity and Autonomy
Cybersecurity is one of the areas where this progression is most pronounced. OpenAI estimates that Astra achieves the "Critical" threshold within its Preparedness Framework—a first for any of its models. With appropriate tools and access, Astra can identify previously unknown vulnerabilities and develop ways to exploit them on highly protected systems without human intervention at every stage. During an internal evaluation, the model notably discovered two zero-day vulnerabilities that were subsequently used in an exploitation chain.
Reinforcing Safety Mechanisms
These advanced capabilities have prompted OpenAI to reinforce protections surrounding the model. The company reports having slowed certain stages of its development and launch in order to improve its mechanisms against malicious cyber usage and unauthorized actions.
Astra integrates features such as built-in refusal logic, security classifiers, action monitoring, and mechanisms capable of interrupting specific operations deemed potentially unauthorized. Furthermore, OpenAI states that Astra adheres more closely to explicit safety restrictions than GPT-5.6 Sol in evaluations. On a set of cybersecurity bypass tests, Astra refused 91.5% of forbidden queries, compared to 59% for GPT-5.6 Sol.
The company acknowledges, however, that monitoring becomes increasingly difficult as model capabilities advance and plans additional controls that may slow down or interrupt legitimate tasks.
The Future of AI and AGI
The arrival of Astra thus comes amid a context where agentic capabilities are simultaneously more useful and harder to control. OpenAI plans a gradual rollout of the model, while access to its most advanced cybersecurity functions remains highly restricted.
Greg Brockman, CEO of OpenAI, estimates that this generation of models could mark entry into a new era of Artificial General Intelligence (AGI), while recognizing that the definition of AGI itself remains subject to interpretation. With GPT-6 Astra, OpenAI crosses a new threshold in agentic model capabilities. However, this advancement also places safety, alignment, and control over autonomous systems at the center of challenges for next-generation artificial intelligence.
