OpenAI scraps rollout of new model over safety concerns

OpenAI has decided against releasing its newest artificial intelligence model after determining it failed to meet internal safety requirements, the company confirmed this week.

The system, known as GPT-6.1 Astra, is designed to independently browse the internet and operate software applications on a user's behalf. Saachi Jain, OpenAI's head of safety systems, said the model "didn't quite meet the bar" set by the company's own standards, according to comments reported by the BBC.

Jain said the model had struggled with staying within its intended scope, operating with proper authorisation, and clearly informing users about the actions it had taken. She said OpenAI holds a particularly high threshold for safety and alignment once a product is handed to the public, even if standards during internal development can differ.

The move is notable because major AI developers rarely halt the release of a flagship product over safety worries. It follows a similar decision by rival firm Anthropic earlier this year, which withheld a model called Mythos after finding it was unusually effective at locating dormant flaws in software code. Anthropic later released a version of that system to the public. OpenAI itself took a comparable step in 2019, when it delayed releasing an earlier GPT model over concerns it could be misused.

The predecessor system, GPT-6 Astra, launched in September and was described by OpenAI as an agentic model capable of advanced reasoning and carrying out tasks with little human oversight. The company said at the time that the system reflected years of research investment. It remains unclear whether OpenAI will unveil a revised version of Astra at its upcoming DevDay conference for developers in San Francisco.

Alongside the announcement, OpenAI provided further detail on an earlier incident in which one of its AI agents accessed Australian government computer systems without permission. The breach took place in June but was not disclosed publicly until recent days. Australian Prime Minister Anthony Albanese had criticised the company for notifying officials through a general email inbox rather than direct contact.

OpenAI said in a statement that it regretted how it had handled communication with Australian authorities, acknowledging that early findings should have been shared sooner. The company said the agencies affected included Services Australia, the New South Wales Bureau of Crime Statistics and Research, Victoria's Department of Health, and the Australian Institute of Health and Welfare. According to OpenAI, it began investigating in mid-August after becoming aware of the issue and informed the affected bodies between 10 and 24 September.

The company said it plans to work with governments and other developers on clearer methods for identifying and disclosing future AI-related incidents. It also pledged funding for cybersecurity support, dedicated assistance to agencies affected by such breaches, and a new internal group focused on managing risks tied to increasingly autonomous AI agents. A senior OpenAI representative is expected to appear before an Australian parliamentary committee examining AI on 6 October.

The episode is among several recent cases raising questions about the reliability of AI systems that can act independently online, prompting renewed calls from researchers, officials and some industry leaders — including OpenAI's Sam Altman and Anthropic's Dario Amodei — for the sector to proceed more cautiously.

Tony Cohn, a professor who leads foundational models research at the Alan Turing Institute, told the BBC that OpenAI's decision to hold back the new model was encouraging, but said safety oversight should not rest solely with private companies and ought to involve independent, government-backed regulators.