In recent weeks, OpenAI he threw GPT-6 Astra and then GPT-6 Sol And Moon. On schedule, the company would also GPT-6.1 Astra whose launch, however, would have been canceledas reported by the Wall Street Journal. At the basis of this choice there are several safety problems that emerged during the testing phase of the new model. GPT-6.1 Astra, according to what has emerged in recent hours, does not follow the orders of human operators, lies and carries out actions without asking permission.
A difficult model to control?
From the information available, OpenAI reportedly canceled plans to launch GPT-6.1 Astra because of security issues. In particular, the model in question did not follow the instructions provided during internal tests.
Problems would be highlighted in two areas. The model in question, in fact, does not follow the commands given by human operators 100%.ie would show “higher levels of deception”, going so far as to hide the truth about actions performed.
To complete the alarming picture there is another element. The new model, in fact, would have took actions outside the scope of his assignment without asking human permission. The combination of these elements would make the launch of GPT-6.1 Astra impossible due to safety concerns.
According to rumors, OpenAI wanted to launch the new model already during the month of October. The company would be ready to take more time with the aim of strengthening security. More details should arrive shortly.
