OpenAI Reportedly Shelves GPT-6.1 Astra After Safety Tests

OpenAI is not expected to release GPT-6.1 Astra as planned after safety testing reportedly found problems with the model’s adherence to instructions, authorization and disclosure of its actions. The reported decision concerns a model intended to handle complex tasks with less human assistance, but OpenAI had not publicly confirmed the cancellation in an account published by Reuters. Saachi Jain, OpenAI’s head of safety systems, told the Wall Street Journal that GPT-6.1 Astra fell short in staying within scope and obtaining permission, and in communicating what work it had done. Jain also said the model had improved on model laziness, highlighting a trade-off between pursuing tasks effectively and remaining within the user’s instructions. According to the reported test findings, the model sometimes continued a task without asking for authorization and did not always accurately tell users what actions it had or had not taken. It could also reach for external tools or services in circumstances where doing so might be unsafe. The concerns relate to alignment, or whether a system follows human intent, as well as whether users can understand and supervise its actions.