The model did not meet safety requirements
The information was first reported by The Wall Street Journal. According to Reuters, the premiere of GPT-6.1 Astra, scheduled for October, was canceled after tests that showed, among other things, problems with compliance with human supervision, scope of authorization and transparency of model activities.
GPT-6.1 Astra was supposed to be the next step in the development of the so-called AI agents. Such systems do more than just answer questions. They can independently browse websites, use applications, search for information or perform multi-step tasks.
Saachi Jain, head of security systems at OpenAI, told the BBC that the new model "didn't quite meet" the company's standards. The problem concerned, among others, whether the model stayed within the permissions granted to it and whether it properly informed the user about what it had done.
This is especially important in the case of autonomous agents. The more activities they can perform without direct human intervention, the more important it is to control their behavior. OpenAI emphasizes that before making the model available to users, it applies very high requirements regarding security and compliance with the assumptions.
Astra already exists. This is about the next version
The name GPT-6.1 Astra can be misleading. OpenAI has already introduced GPT-6 Astra. The company announced this model on September 3, 2026, presenting it as a system intended, among others, for: for programming, research, computer use and performing complex, multi-step tasks.
The planned GPT-6.1 Astra was supposed to be its next version. Its release has now been suspended. The decision was made the day before OpenAI's annual DevDay developer conference in San Francisco. The company will present new solutions during the event. It is not known whether it will present a different version of Astra or a changed development schedule.
AI has gained access to government systems
OpenAI's decision came at a time of growing pressure on the company over the behavior of its AI agents. One of the most serious cases is the incident in Australia. In June, an OpenAI agent hacked into the government's health care system during a task related to searching for information on drug spending.
According to the Australian government, the agent had access to both public and non-public files. At this stage, there is no information about access to patients' personal data. An investigation is underway.
Australian authorities also reported that the model interacted with four public websites in June. These were: Services Australia, the Australian Institute of Health and Welfare, the Victoria Department of Health and the New South Wales Office of Crime Statistics and Research.
According to the Australian government, OpenAI became aware of the incident in August. Services Australia was only notified on September 10, via a general email address used, among others, to report security issues.
Australian Prime Minister Anthony Albanese said that the agent had found a way to bypass the website's security. He emphasized that the attacked system contained primarily public statistical data and was not a system handling individual cases or Medicare payments.
OpenAI apologized for the incident and announced additional actions related to cybersecurity and risk management resulting from increasingly autonomous agents.
