OpenAI shelves plans to release new model over safety concerns

OpenAI has shelved plans to release its latest AI model over safety concerns.  

The system, known as GPT-6.1 Astra, "didn't quite meet the bar" of the company's standards, according to OpenAI's head of safety systems, Saachi Jain.

The model was designed to browse the internet and operate apps by itself. However, internal tests raised questions about whether the AI would follow users’ instructions.

"For anything regarding safety and alignment, there's a trade off," said Jain. "You really do need to find what's the right line between staying within scope, but also avoiding laziness in terms of how the model actually pursues tasks even when it hits friction."

Jain added: "While [GPT-6.1 Astra] improved on axes such as model laziness, it didn't quite meet the bar in terms of staying within scope and authorisation, and how it communicates back to the user about the type of work it's done."

This is a rare example of one of the big AI companies cancelling a new release over safety concerns.  

Anthropic made the decision earlier this year not to release a Claude model, named Mythos, because it was too good at finding dormant software bugs. However, months later it released a version of model to the public.

The decision by OpenAI comes after news last week that an agent belonging to the company hacked into an Australian Government website in June and accessed private data.

This is the first publicly reported case of an AI model breaching a government system.  

OpenAI said it became aware if the breach in August when it reviewed "misaligned model activity" and contacted a general Australian Government agency email address on 10 September.