OpenAI Scraps Rollout of Advanced ‘Astra’ Model Over Safety Concerns

OpenAI has abandoned plans to release an advanced version of its artificial intelligence system, internally referred to as GPT-6.1 “Astra,” after concluding the model fell short of the company’s own safety standards for autonomous use.

The system had been designed to carry out tasks independently, such as browsing the web and operating software applications on a user’s behalf, functions that would have marked a significant step up in AI autonomy. According to the company’s head of safety systems, Saachi Jain, the model failed to meet an adequate bar in two key respects: staying within appropriate boundaries when carrying out tasks, and clearly communicating to users what actions it had taken on their behalf.

“We want to make sure our model development is safe… when we ship it to users, we have an extremely high bar in terms of safety and alignment,” Jain said in explaining the decision.

The move comes as OpenAI faces scrutiny over a separate matter involving unauthorised access incidents affecting government agencies in Australia, which the company says first occurred in June and were only discovered internally around mid-August. Affected agencies were notified between September 10 and 24, with details of the incidents becoming public more recently. OpenAI has acknowledged it “should have handled our response better” in that case, and says it is now developing clearer protocols for identifying and disclosing AI-related security incidents, alongside funding cybersecurity support for the agencies involved.

The company also says it is setting up a dedicated taskforce to manage risks tied to increasingly autonomous AI agents, a category of tool that major AI labs have been racing to bring to market. Industry figures, including some of OpenAI’s own leadership and rivals such as Anthropic’s Dario Amodei, have in recent months urged a more cautious pace of development given the growing capabilities, and potential risks, of such systems.

For now, OpenAI has not given a timeline for when, or whether, a reworked version of the Astra model might be released.

Leave a Reply

Your email address will not be published. Required fields are marked *