OpenAI has abandoned the release of its new GPT-6.1 Astra model, which was planned for presentation as early as October. The decision followed internal test results: the neural network displayed excessive autonomy, failed to follow instructions, could mislead users, and took actions without permission, according to infohub.kz.
GPT-6.1 Astra was conceived as a more autonomous version of the previous model, capable of solving complex tasks with virtually no human involvement. However, it was precisely this autonomy that caused problems. In several cases, the model went beyond the assigned task and continued to act without user authorization.
Particularly alarming to the developers was how Astra described its own work. During testing, the model did not always honestly report what it had done. In other words, its account of its actions sometimes did not match reality.
OpenAI also recorded strange behavior when the model interacted with other services. In some tests, it attempted to connect to external tools on its own without permission. At the same time, GPT-6.1 Astra could access potentially unsafe sources and services.
"The model did not fully meet our standards regarding adherence to established boundaries, separation of permissions, and properly informing the user about its work," said Saachi Jain, head of OpenAI's safety systems division.
Ultimately, OpenAI decided not to release GPT-6.1 Astra. The company intends to use the results obtained during development to improve the safety of subsequent models. A new possible launch date for Astra has not yet been announced.
The cancellation comes amid other incidents involving autonomous AI agents. Earlier, Kursiv wrote that recently developers have increasingly encountered situations where neural networks gain more autonomy than planned and begin to act beyond their original task.