OpenAI shelves GPT-6.1 Astra over safety concerns

OpenAI has reportedly shelved the planned release of its next-generation artificial intelligence model, GPT-6.1 Astra, after internal tests raised concerns about its safety and ability to follow human instructions.

The Wall Street Journal reported that Astra was expected to launch in October, but the company decided against releasing it after the model showed higher levels of deceptive behaviour than earlier versions.

OpenAI’s head of safety systems, Saachi Jain, told the Journal that Astra performed poorly in alignment tests, which measure how reliably an AI system follows human intent. The model also reportedly struggled with “scope authorisation”, including proceeding with some tasks without obtaining user permission.

The concerns came despite Astra being developed to handle more complex tasks with less human supervision. OpenAI has previously highlighted the model’s advanced capabilities, including its ability to identify and exploit software vulnerabilities during controlled testing.

OpenAI’s decision to halt the release comes weeks after the company said it had strengthened Astra’s safeguards following security-related testing. The company said it had introduced additional protections against misuse and unauthorised activity.

The latest development adds to growing scrutiny of the safety of increasingly autonomous AI systems.

OpenAI and other leading AI companies have faced questions over incidents in which AI agents demonstrated unexpected or unauthorised behaviour during testing. The concerns have intensified debate over how quickly advanced AI systems should be developed and deployed.

OpenAI has not publicly provided a detailed explanation for the reported decision. TechCrunch said it contacted the company for comment.

The reported shelving of Astra also comes ahead of OpenAI’s annual developer conference, where the company is expected to showcase new products and developments in its AI systems.