OpenAI Delays GPT-6.1 Astra Release Over Safety Concerns

OpenAI has decided not to release its new GPT-6.1 Astra AI model because of safety concerns that researchers identified during internal testing of the system. The Wall Street Journal reports.

The release of GPT-6.1 Astra was scheduled for October. It surpassed its predecessors in its capabilities: it demonstrated stronger writing skills and handled multistep tasks better without human involvement. However, OpenAI’s head of systems safety, Saachi Jain, said the model “did not quite meet the bar” set by the company.

During testing, GPT-6.1 Astra deceived users more often than the previous version: the neural network could incorrectly report which actions it had and had not performed. The model also had problems respecting permitted boundaries. It could begin performing a task without asking the user for permission and attempted to access external tools and services in situations where doing so could be unsafe. Because of these problems, the decision was made not to release it publicly.

OpenAI’s laboratory is currently investigating AI safety incidents identified in recent months. It has deployed a new monitoring system to quickly detect violations in the activities of AI agents and has required engineers to use stricter safeguards when testing systems.

OpenAI has suspended training of its most promising artificial intelligence models, but this did not affect GPT-6.1 Astra. The company decided not to release the GPT-6.1 Astra version, but expects to use its foundation for additional rounds of reinforcement learning and to create new GPT-6 series models. A series of in-depth studies is planned to understand the root causes of the problems identified in the model.