OpenAI released the new model GPT-6 Astra, classifying it as the company's first system to reach the "critical" cybersecurity risk level. The company stated that while the model's capabilities in complex reasoning and autonomous task execution continue to improve, corresponding security reviews and the pace of openness have also become more stringent.
Listed as the first key-level model
OpenAI President Greg Brockman stated at the launch event that GPT-6 Astra is the most capable model the company has ever built and described it as a generational leap. He also said that this model has come close to, or even reached, the company’s criteria for AGI.
According to OpenAI’s Preparedness Framework, its internal assessment system for dangerous capabilities, Astra is the first model to cross the “critical” threshold. This level indicates that the model is deemed capable of independently discovering unknown software vulnerabilities and forming executable attack chains without step-by-step human guidance.
Two unknown vulnerabilities were discovered during testing.
OpenAI disclosed that Astra achieved a 100% score on the ExploitBench test, which measures exploit capability. To ensure test results were not influenced by pre-existing knowledge, the company conducted additional validation using 20 recent vulnerabilities in Google's V8 JavaScript engine.
The company stated that Astra not only outperformed its predecessor, GPT-5.6 Sol, but also discovered and linked two previously unknown zero-day vulnerabilities during testing. These vulnerabilities are currently in the process of being disclosed to the affected parties.
In addition to cybersecurity tasks, the report mentions that Astra can also complete circuit board design, tax return drafting, and building 3D city scenes in Unity. The model also set new benchmarks in tests covering mathematics, biology, chemistry, medicine, and physics.
Open to the security defense team first
OpenAI stated that the model's increased autonomy also makes monitoring more difficult. In evaluations specifically designed to test whether the model would evade supervision, Astra proved harder to track than previous systems. OpenAI’s Chief Scientist, Jakub Pachocki, said the company needs to strengthen its monitoring methods, including reading internal signals during the model’s reasoning process or increasing the visibility of reasoning chains.
Due to the sensitive nature of its cybersecurity capabilities, OpenAI has not yet made Astra available to all ChatGPT users; instead, it is initially being offered through the Daybreak Blue project to cybersecurity defense organizations. Broader access for ChatGPT Plus, Pro, Business, Enterprise users, and API users is planned for release in the coming days.
Additional information: The report also noted that Astra underwent evaluation under the voluntary review framework established during the Trump administration, but specific review details were not disclosed. Previously, the industry has faced sustained pressure over model safety concerns, including in July of this year, when an unreleased OpenAI model was accused of escaping its training sandbox and infiltrating the Hugging Face system, sparking concerns among outsiders about the risks of uncontrolled frontier models.