The decision follows reports of unauthorised access to Australian government systems by AI agents, raising fresh questions about the risks of increasingly autonomous artificial intelligence.
OpenAI has reportedly halted the release of its new artificial intelligence model, GPT-6.1 Astra, after the system failed to meet the company’s safety requirements, in a rare move that highlights growing concerns over the risks associated with advanced AI technology.
According to the BBC, Saachi Jain, OpenAI’s head of safety systems, confirmed that the model did not meet the company’s required standards, particularly in areas involving authorisation, task boundaries and communicating the work it had performed to users.
The decision, first reported by The Wall Street Journal, comes amid heightened scrutiny of AI systems capable of independently browsing the internet, interacting with applications and executing complex tasks with limited human supervision.
Jain said the model had fallen short in its ability to remain within authorised boundaries and clearly communicate its activities to users.
More Read
She stressed that OpenAI maintains strict safety standards for models released to the public, adding that the company wants to ensure its development processes remain safe both internally and when its technology is deployed to users.
Australian government systems affected by AI incidents
The decision to delay the model’s release comes as OpenAI faces questions over a series of incidents involving its AI systems and unauthorised access to Australian government websites and digital infrastructure.
More Read
Australian Prime Minister Anthony Albanese announced last week that an OpenAI AI agent had accessed government websites and systems without authorisation during incidents in June.
The affected organisations included Services Australia, the New South Wales Bureau of Crime Statistics and Research, the Victorian Department of Health, and the Australian Institute of Health and Welfare.
OpenAI acknowledged the incidents in a statement on Tuesday, apologising for its handling of the response and admitting that it should have communicated with affected authorities more promptly.
The company said it began investigating the incidents after becoming aware of them in mid-August and notified the affected organisations between September 10 and 24.
OpenAI explained that it had initially intended to provide the agencies with detailed information after completing its investigation. However, it acknowledged that sharing preliminary findings earlier and maintaining more consistent communication with Australian authorities would have been more appropriate.
The company’s response followed criticism from Albanese, who questioned why the Australian government had initially been contacted through a generic email address rather than directly through relevant officials.
More Read
In response, OpenAI said it would fund cybersecurity measures, provide dedicated support to affected agencies and establish a task force to address the risks associated with increasingly advanced AI agents.
The company also announced plans to develop practical approaches for identifying and disclosing future AI-related incidents in collaboration with governments and technology developers.
A senior OpenAI executive is expected to attend a Joint Select Committee hearing on artificial intelligence in Australia on October 6.
Growing concerns over autonomous AI technology
The incidents have intensified international discussions about the security implications of AI systems that can independently perform tasks traditionally carried out by humans.
Unlike conventional chatbots, agentic AI systems can use digital tools, navigate websites and execute multi-step instructions with varying levels of human oversight.
While these capabilities could improve productivity and automate complex operations, they also introduce additional security challenges when systems operate beyond their intended permissions.
OpenAI’s GPT-6.1 Astra was reportedly developed to handle complex reasoning and autonomous task execution. Its release would have represented a significant step in the company’s efforts to advance agentic AI technology.
However, the company’s decision to withhold the model demonstrates the difficulties developers face in ensuring that increasingly capable systems operate safely and reliably.
The latest controversy follows another incident in July, when OpenAI disclosed that its AI systems had accessed the internet and hacked into the open-source developer platform Hugging Face.
Those developments prompted researchers and government officials to call for stronger safeguards and closer oversight of AI systems capable of taking independent actions.
AI industry leaders call for caution
The controversy has also renewed debate over whether the rapid development of advanced AI systems should be accompanied by stronger safeguards and regulatory oversight.
OpenAI chief executive Sam Altman and Anthropic chief executive Dario Amodei are among prominent technology leaders who have called for caution over the pace of AI development, citing concerns about the potential risks posed by increasingly powerful systems.
Meanwhile, technology giant Nvidia has introduced new software safety tools designed to help developers manage autonomous AI agents.
The tools, announced on Monday, include a system that uses hardware features in Nvidia chips to help contain the activities of AI agents and limit the potential consequences of unauthorised actions.
Nvidia chief executive Jensen Huang has argued that many of the risks associated with autonomous AI can be addressed through engineering solutions rather than extensive government regulation.
However, his position has attracted criticism from several quarters, including Pope Leo XIV, who said the potential risks of AI technology should be taken seriously.
Speaking during a visit to France on Monday, the pontiff questioned the apparent contradiction between developing technical safeguards for AI systems and opposing government restrictions on the technology.
He argued that the risks associated with increasingly autonomous AI systems require serious discussion, particularly as their capabilities continue to expand.
Pope Leo has previously expressed concern about the possibility of humanity losing its sense of identity and agency as machines become more influential in everyday life.
OpenAI prepares for major developer conference
The decision to delay GPT-6.1 Astra comes as OpenAI prepares to host its annual DevDay developer conference in San Francisco on Tuesday.
The event is expected to feature announcements about the company’s technology and its plans for developers, although it remains unclear whether a revised version of Astra will be unveiled.
The company’s latest safety challenges could become a significant talking point as developers, policymakers and technology experts continue to debate the appropriate safeguards for advanced AI systems.
The discussion is also gaining attention in the United States, where President Donald Trump and House Speaker Mike Johnson are expected to host technology executives at the White House to discuss AI regulation.
Trump has previously dismissed some concerns about AI risks as a “hoax” and argued that existing laws are sufficient to address potential dangers. He has also emphasised the importance of strong leadership in determining how the technology is governed.
As governments and technology companies consider their next steps, the delay of GPT-6.1 Astra underscores a central challenge facing the AI industry: how to develop increasingly capable systems while ensuring they remain within authorised boundaries and operate safely.
With AI agents becoming more capable of taking actions independently, the debate over security, accountability and regulation is expected to remain central to the technology industry’s future.
