Sudden! OpenAI next-generation model Astra project changes

Compiled by Zhidongzhi | Edited by Yang Jingli | Li Shuiqing Zhidongzhi reported on August 19 that early this morning, OpenAI announced that the company has temporarily slowed down the development pace of cutting-edge models including Astra, and a large number of Astra training and evaluation tasks are still suspended. OpenAI once suspended reinforcement learning (RL) training of the latest model for deployment for two weeks. During this period, it further strengthened the research environment, conducted red team testing, and expanded the coverage of the monitoring system. ▲OpenAI temporarily slows down the pace of cutting-edge model research and development. Currently, the largest-scale cutting-edge model reinforcement learning training originally planned by OpenAI has not yet resumed. The company is observing model behavior, verifying safety measures, and accumulating more model alignment evidence through smaller-scale training and evaluation. OpenAI CEO Sam Altman subsequently issued a statement stating that the suspension of reinforcement learning training for some cutting-edge models is to ensure that model alignment, security and monitoring measures can adapt to the latest capability level of the model. He believes that the entire AI industry ultimately needs to collaborate around unified safety standards. Before a unified standard is formed, OpenAI will take action on its own. In the future, confidence in model safety will increasingly determine the speed of AI development. ▲ Altman issued a response. On August 2, OpenAI officially announced the next-generation model Astra and announced 10 new achievements in mathematics and theoretical computer science achieved by its internal version. However, within a few weeks, multiple models, including GPT-5.6 Sol and a stronger pre-release model, exploited vulnerabilities to break through the isolation environment during internal network security evaluations, then invaded the Hugging Face system and obtained evaluation answers from the production database. In addition, preliminary assessments indicate that Astra may have reached the "critical level" of cybersecurity capabilities defined in the Readiness Framework, thus suspending some training and evaluation tasks. The so-called "critical level" network security capabilities refer to the model's ability to attack hardened critical systems without manual intervention, or to autonomously plan and execute end-to-end network attacks. OpenAI has not yet been confirmed