The problems were discovered during an internal test of the new model "GPT-6.1 Astra", which was planned for release in October.
“We want to ensure that our model development is secure whether it happens internally within the company, or when we deliver it to users,” said Saachi Jain, head of security systems at OpenAI, in a statement, according to The Wall Street Journal.
One of the model's shortcomings is reported to be compliance, that is, how well it follows what the user wants it to do. Furthermore, problems were discovered where the model proceeded with a task without asking the user for permission and, in some cases, potentially unsafe external tools and services were used.
