News on August 19,OpenAI ACCORDING TO THE COMMUNICATION, THE COMPANY HAD PREVIOUSLY SUSPENDED TWO WEEKS OF INTENSIVE LEARNING (RL) TRAINING ON THE LATEST MODEL IN ORDER TO STRENGTHEN THE RESEARCH ENVIRONMENT, CONDUCT RED TEAM TESTING AND EXPAND SURVEILLANCE COVERAGE. CURRENTLY, SOME OF THE TRAINING HAS RESUMED, BUT THE MOST PLANNED FORWARD MODEL RL TRAINING REMAINS SUSPENDED。
This adjustment is mainly related to two developments: the recent OpenAI and Hugging Face models, which assess security incidents; and the preliminary assessment, which indicates that the Astra model, which has not yet been released, is likely to meet the "Key Cybersecurity Capacity" threshold defined in the RPF。
OpenAI has strengthened the code and network segregation of the training environment and expanded real-time monitoring of model behaviour. The monitoring system will check the call and complete operation of the tool and identify risks such as unauthorized access, data theft and disruptive operations. Astra and other high-capacity models are subject to monitoring when using tools for reasoning, training or evaluation。
OpenAI indicated that it would verify security measures through small-scale training and evaluation, build up more matching evidence before deciding whether to resume the most extensive training. CEO Sam Altman also indicated that the company was still planning to launch a new model in the near future, and that this suspension would only affect the longer-term distribution plan。