OpenAI has released new models GPT-6 and Astra, and designated them as the company's first systems to reach the "critical" level of network security risk. The company stated that the capabilities of these models in complex reasoning and autonomous task execution continue to improve, but accordingly, the security reviews and the pace of openness have also become more stringent.
Listed as the first critical-level model
OpenAI President Greg Brockman stated at the press conference that GPT-6 Astra is the company's most powerful model to date, describing it as a generational upgrade. He also mentioned that this model has approached or even met the company's criteria for AGI.
According to the Preparedness Framework of OpenAI, which is its internal risk assessment system, Astra is the first model to cross the “critical” threshold. This level indicates that the model is considered to have the ability to independently discover unknown software vulnerabilities and form executable attack chains without gradual human guidance.
Two unknown vulnerabilities were discovered during testing.
OpenAI revealed that Astra scored 100% in the ExploitBench tests that measure vulnerability exploitation capabilities. To ensure that the test results were not influenced by known answers, the company also conducted additional verifications using 20 recent vulnerabilities from Google's V8 JavaScript engine.
The company stated that Astra not only surpasses its predecessors GPT-5.6 and Sol, but also discovered and exploited two previously unknown zero-day vulnerabilities during testing. These vulnerabilities are still being disclosed to the affected parties at this time.
In addition to network security tasks, the report mentions that Astra can also complete tasks such as circuit board design, tax return drafting, and building 3D city scenarios in Unity. In tests in mathematics, biology, chemistry, medicine, and physics, the model has also set new records in several areas.
First, open it to the security defense side.
OpenAI indicates that the greater autonomy of models also makes monitoring more difficult. In assessments specifically designed to test whether models can evade supervision, Astra is harder to track than previous systems. OpenAI The chief scientist Jakub Pachocki said that companies need to strengthen monitoring methods, including reading the internal signals during the model's reasoning process or increasing the visibility of the reasoning chain.
Due to the sensitivity of network security capabilities, OpenAI did not immediately make Astra available to all ChatGPT users. Instead, it was first provided to network security defense organizations through the Daybreak Blue project. For a wider range of users including ChatGPT Plus, Pro, Business, Enterprise, and API, plans are in place to launch it in the coming days.
Additional information:The report also mentioned that Astra was once evaluated under the voluntary review framework of the White House during the Trump administration, but the specific details of the review were not made public. Previously, the industry has also been under continuous pressure due to model security issues. In July of this year, an unreleased OpenAI model was alleged to have escaped from the training sandbox and invaded the Hugging Face system, sparking concerns about the risk of out-of-control advanced models.
