August 8 News.OpenAI Today's statement states that internal assessments indicate that research is under wayModel Astra Significant progress has been made on "Intelligent Encoding and Cybersecurity" and, in combination with expert assessment, the company "can't rule out" It has the potential to reach the "significant" Internet security threshold in the Preparadness Platform。
OpenAI immediately suspended Astra ' s internal activities that did not meet the new security requirements and introduced stricter security controls for the higher-capacity model。
According to the OpenAI definition, the `significant' threshold is that the model can, without intervention, identify and use functional zero-day holes in multiple high-protected critical systems, or simply a high-protected target, design and implement a new end-to-end attack strategy. Before, GPT-5.6 Sol and other models were rated "High"。
OpenAI states that Astra has nothing to do with the Hugging Face invasion in July。
Controls for Astra include: isolation testing environment, restricted access to networks and tools, sandbox implementation, enhanced model weight protection and encryption, and general control of all smart application "high-risk behaviour and malpractice " ; follow-up testing will also be conducted in cooperation with relevant government agencies and AI security organizations。
Previously, the implosioner @synthwavedd stated that OpenAI was preparing to launch a new model called `mewfour ' , Astra, the internal target time is next week. He described Astra as the largest new pre-training model for OpenAI after GPT-4.5。