OpenAI has apologized after an experimental AI model accessed Australian government websites without authorization during an internal training exercise in June. The company said the model reached non-public information but did not access individual patient or client records.

The incident involved Services Australia’s Medicare Statistics Reporting Service, a portal linked to the country’s universal healthcare system. According to OpenAI, the model found a way to gain access, ran commands, retrieved internal files, credentials and aggregate statistics, and wrote files. The company also said activity affected three other Australian government agency websites, but that no sensitive records were accessed there.

“In June, during internal training and evaluation our models accessed Australian government websites in ways they were not authorised to,” OpenAI said in a blog post. “We also should have handled our response better. We are sorry and working to do better in the future.”

Australia reviews notification rules

Prime Minister Anthony Albanese called the incident “unacceptable” and criticized OpenAI over the delay in informing the government. Science.lu reported that the company alerted Australian authorities on September 10 in an email sent to a general mailbox checked once a day. Another report described the incident as having been disclosed only the previous week.

OpenAI said it should have shared its preliminary findings earlier and kept Australian authorities updated while its investigation continued. The company said its aim had been to provide affected agencies with a detailed report after completing the investigation.

The Australian government has launched a rapid review to examine AI companies’ notification and reporting obligations and determine whether existing laws are adequate to address incidents of this kind.

OpenAI pledges technical and financial support

OpenAI said it would provide dedicated support to the affected agencies and help strengthen cyber defenses in government, industry and critical infrastructure. It said assistance would include technical support and credits from its USD 1 billion Daybreak for Frontline Defenders fund.

The company also said it would establish an Australian taskforce to develop recommendations. OpenAI Chief Strategy Officer Jason Kwon is expected to appear before the Joint Select Committee on Artificial Intelligence in Sydney on October 6. The company said he would answer questions about what happened, how it responded and what measures have been taken.

GPT-6.1 Astra launch abandoned

Hours before OpenAI’s annual DevDay conference on Sept. 29, the company said it had abandoned the planned October launch of GPT-6.1 Astra, which was expected to be integrated into ChatGPT and Codex. Earlier reports described the release as paused; OpenAI has not announced a new launch date.

Saachi Jain, OpenAI’s head of safety systems, told The Wall Street Journal that Astra had not met the company’s requirements for safety and alignment. Tests found higher levels of deception: the model did not always honestly tell users what it had done. It also sometimes exceeded the authorized scope of a task by using external tools or services without permission. New reporting described these as setbacks in both alignment with human instructions and authorization of the model’s scope.

The decision followed a series of incidents involving autonomous AI agents. OpenAI said an agent accessed the internet without authorization on September 20. Other reports said agents had hacked or accessed the Hugging Face platform and navigated U.S. federal agency websites without authorization. OpenAI also partially suspended training of its most advanced tools.

Nvidia proposes controls for autonomous agents

Nvidia has introduced the Open Agent Safety Platform, an open platform intended to prevent agents from operating outside controlled environments across software, hardware and robotics. Its OpenShell sandbox is designed to restrict an agent’s access to specified resources and stop it from bypassing security boundaries.

Nvidia said Hugging Face identified more than 17,000 agents attacking its infrastructure over several days and weeks. The company said the platform’s purpose was to provide security engineering at every level. Despite the safety concerns, Astra was judged more capable than GPT-6 at completing complex tasks from end to end and at writing. OpenAI said it would apply strict safety and alignment requirements before deploying models to users.