OpenAI has cancelled the planned release of its GPT-6.1 Astra artificial intelligence model after internal tests found problems with deception, user authorisation and compliance with instructions.
The company had planned to launch the model in October and integrate it into ChatGPT and its Codex coding tools. GPT-6.1 Astra was designed to complete complex tasks with less human involvement.
The decision marks a change in OpenAI’s release plans as AI developers face scrutiny over systems that operate across websites, applications and other external services.
Saachi Jain, OpenAI’s head of safety systems, told The Wall Street Journal that GPT-6.1 Astra failed to meet the company’s standards in two areas.
The model performed worse than its GPT-6 Astra predecessor in alignment tests measuring whether a system follows human instructions and intended goals.
Testing also found more deceptive behaviour. In some cases, the model failed to give users an accurate account of actions it had taken or omitted.
Astra also continued some tasks without seeking permission and attempted to access external tools or services in situations involving safety risks, according to reports about the tests.
“It didn’t quite meet the bar in terms of staying within scope and authorization, and how it communicates back to the user about the type of work it’s done,” Jain said, according to Reuters.
She said OpenAI applies a high safety and alignment threshold before releasing models to users.
The decision follows scrutiny of OpenAI systems after incidents involving Australian government networks.
Australian Prime Minister Anthony Albanese said last week that an OpenAI agent had accessed government websites and systems without authorisation in June. OpenAI said on Tuesday that Services Australia, the NSW Bureau of Crime Statistics and Research, the Victorian Department of Health and the Australian Institute of Health and Welfare were affected.
The company apologised for its handling of the incidents and said it should have informed Australian authorities sooner. OpenAI said it would fund cyber security measures, provide support to affected agencies and establish a taskforce focused on risks from AI agents.
The UK AI Security Institute has also tested GPT-6 Astra. The institute found the model carried out unauthorised attack activities in simulations more often than earlier OpenAI models.
OpenAI chief executive Sam Altman and Anthropic chief executive Dario Amodei have both backed calls this month for a slower pace of AI development and stronger safeguards.
Anthropic has previously withheld a Claude model over concerns about its ability to identify software vulnerabilities before releasing a version months later.
HT
