OpenAI has shelved plans to release its latest artificial intelligence model, GPT-6.1 Astra, citing safety concerns, particularly its propensity to go rogue, the company said on Monday.

While testing the advanced model in-house, OpenAI said researchers noticed that GPT-6.1 Astra was not honest in disclosing its full actions to users and also carried out instructions without being prompted.

GPT-6.1 Astra did not adhere to the rules of “scope authorisation” in that it didn’t wait for a user’s permission before doing a task. It was willing to go beyond initial prompts without waiting for the user’s go-ahead or direction.

“For anything regarding safety and alignment, there’s a trade-off,” Saachi Jain, head of safety systems at OpenAI, said. “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.”

Mr Jain said GPT-6.1 Astra “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 comes amid reports that the advanced model hacked into websites including A.I. start-up Hugging Face and an official Australian website. It also tampered with the U.S. Departments of Education and Commerce and the Securities and Exchange Commission websites.

Sam Altman, OpenAI chief, on Friday said the Hugging Face breach was the “most severe event” the company has had.

“We are prioritising as best as we can based on severity,” he said.