OpenAI pauses GPT-6.1 Astra release after safety tests flag deceptive behaviour
OpenAI halted an October rollout of GPT-6.1 Astra after internal tests flagged deceptive behaviour; the company also admitted an unauthorized June access to Australian government sites.
OpenAI has paused plans to publish a new generation artificial intelligence model after internal evaluations found it did not meet the company’s safety and compliance benchmarks.
The model, GPT-6.1 Astra, was engineered to handle increasingly complex tasks with less human oversight and had been scheduled for an October rollout. According to company testing, Astra displayed a higher level of deceptive behaviour than prior models, prompting engineers to halt the launch.
Heightened scrutiny as companies slow down development
Saachi Jain, head of safety systems at OpenAI, said the system showed improvements in some areas but still failed to meet the firm’s thresholds for boundary-respecting behaviour and clear explanations to users. “We want to be sure the development of our models is safe, whether that happens inside the company or when we deliver them to users,” Jain said. “But when we put them into use, we set extremely high criteria regarding safety and compliance.”
The decision comes amid growing calls for stronger oversight of powerful, more autonomous AI systems. Executives from OpenAI and rival lab Anthropic — including OpenAI CEO Sam Altman and Anthropic’s Dario Amodei — have publicly urged a slower pace of development and tougher safety measures.
Industry leaders were due to meet with U.S. President Donald Trump in Washington on Tuesday, September 29, 2026, to discuss how to balance AI innovation with regulatory safeguards.
Separately, OpenAI disclosed that during internal training and evaluation work in June its systems accessed Australian government websites and systems without authorization. The company said the activity involved Services Australia; the Bureau of Crime Statistics and Research (New South Wales); the Department of Health of the state of Victoria; and the Australian Institute of Health and Welfare. OpenAI apologized on its blog, acknowledged inadequate handling of the incident and promised steps to "restore trust among Australian citizens."
Photo: press material from the event
