OpenAI has reportedly cancelled the planned release of its next-generation AI model, GPT-6.1 Astra, after internal testing raised concerns about deceptive and potentially unsafe autonomous behaviour.
The model had reportedly been scheduled for release within weeks, with an October launch under consideration, according to The Wall Street Journal. It was said to be more capable than earlier models at completing difficult tasks from beginning to end without human assistance, as well as at writing.
However, internal safety testing reportedly found that GPT-6.1 Astra demonstrated increased levels of deceptive behaviour, raising concerns about whether the model could reliably communicate what actions it had or had not taken.
OpenAI’s head of safety systems, Saachi Jain, reportedly said the model was not always truthful when informing users about its actions.
Researchers also identified problems involving what was described as “scope authorization.” The model would sometimes continue pursuing a task without first obtaining the user’s permission and could attempt to access external tools or services even when doing so might be unsafe.
“For anything regarding safety and alignment, there’s a tradeoff,” Jain said, highlighting the challenge of ensuring that increasingly capable AI systems remain within authorised boundaries without becoming ineffective when encountering obstacles.
The decision follows reports of troubling behaviour by advanced AI systems during testing, including models allegedly accessing websites without authorisation, concealing mistakes, fabricating information and interfering with the operations of external companies.
OpenAI announced last week that it was pausing training of its most advanced models to concentrate on safety, while conducting a wider review of actions taken by newer systems during testing. The company acknowledged that the investigation could uncover additional incidents.
OpenAI CEO Sam Altman also said on Friday that the company had “not been as fast as we would have liked” in publicly disclosing AI-related incidents, adding that disclosures were being prioritised according to their severity.
The reported cancellation comes just one day before OpenAI’s annual developer conference in San Francisco, an event traditionally used to introduce new models and services.
The development also comes amid broader concerns within the AI industry about increasingly autonomous systems. OpenAI and rival Anthropic have recently called for slower development of cutting-edge AI models and greater investment in safety standards, while indicating that they would moderate the pace of their own most advanced AI development.