OpenAI has unveiled its latest and most sophisticated flagship AI system, GPT-6 Astra. The launch is part of the company’s larger push toward artificial general intelligence (AGI) which is AI that can learn, reason and handle a wide array of intellectual tasks associated with humans.
GPT-6 Astra targets challenging use cases such as advanced reasoning, software development, research, and autonomous agent workflows.
GPT-6 Astra Introduces New Cybersecurity and Safety Measures #
As per OpenAI, GPT-6 Astra has been able to showcase its capability of identifying some hitherto unknown security vulnerabilities and developing attacks on more than one highly secured systems. It has claimed that this model has been made to deal with some very complicated end-to-end tasks such as deep reasoning, coding, software engineering, computing interaction, research and document generation.
The users may choose various reasoning capabilities starting from the lower end up to the maximum end while using the API of the model. Moreover, the model accepts not only textual inputs but also image-based inputs along with the function calling, structured responses and streaming.
This is GPT-6 Astra.
— OpenAI (@OpenAI) Anything you can do on a computer, Astra can do for you. Fast.[pic.twitter.com/gDd0IsewJw][September 3, 2026]
It should be mentioned here that GPT-6 Astra has become the first model to reach the level of critical as per the Preparedness Framework of the company.
Internal testing found that GPT-6 Astra achieved a perfect score of 100 percent on ExploitBench.
Another test revealed that the model was able to develop an attack chain using the browser to escape the sandbox environment and run commands on the host system. Another test had Astra developing a local privilege escalation attack by exploiting multiple vulnerabilities.
With the cyber abilities of the model being upgraded, OpenAI has also put in some additional security measures. According to the company, Astra is much more difficult to jailbreak than GPT-5.6 Sol and it blocks 91.5 percent of the cyber-based jailbreak attempts. In addition, the firm found about half the number of high severity alignment problems in a simulation comprising 54,000 internal Codex tasks.
Also See: iPhone Ultra Could Feature MagSafe Charging Without Sacrificing Slim Design