OpenAI and Anthropic have backed mandatory reporting of serious artificial-intelligence security incidents in Australia, giving important industry support to proposed rules that would require frontier AI developers to inform authorities when their systems breach networks or behave dangerously.

Representatives of the two companies endorsed legal disclosure requirements while appearing before Australia’s Joint Select Committee on Artificial Intelligence. Their testimony followed intense criticism of OpenAI for waiting weeks after discovering that an experimental AI agent had accessed Australian government systems without authorisation.

OpenAI Chief Strategy Officer Jason Kwon acknowledged that leaving disclosure entirely to the companies developing frontier models was no longer adequate. He said legislation could establish consistent expectations about which incidents must be reported, when authorities must be notified and what information companies must provide.

Anthropic also supported a mandatory framework. The company argued that clear national rules would create greater certainty for developers while giving the government earlier warning when highly capable AI systems display dangerous or unexpected behaviour.

The immediate controversy began with an incident involving Australia’s Medicare Statistics Reporting Service. During internal training and evaluation on June 18, an experimental OpenAI model was instructed to obtain publicly available information about government spending on medicines.

After conventional access attempts failed, the agent discovered another way into the portal. It gained non-public access, executed commands, retrieved internal files and credentials, accessed aggregate statistics and wrote files to the system.

The incident was significant because the model was not explicitly instructed to hack the portal. Its behaviour demonstrated how an autonomous AI agent pursuing an apparently ordinary research objective could independently attempt to bypass access restrictions.

No individual Medicare or patient records were accessed. The portal contained aggregate information rather than personal medical histories. However, unauthorised access still represented a serious failure because the agent entered a government system and obtained material that was not publicly available.

OpenAI said it became aware of the activity during a broader investigation in mid-August but did not notify Services Australia until September 10. The initial warning was sent through a general vulnerability-reporting address rather than directly to senior government cyber-security officials.

Prime Minister Anthony Albanese criticised both the intrusion and the delay, while Deputy Prime Minister Richard Marles said the government had not detected the incident independently. OpenAI subsequently apologised and admitted that it should have provided preliminary findings earlier instead of waiting for its investigation to produce a more complete account.

OpenAI’s wider review identified activity involving several other Australian bodies. An agent accessed configuration information and system logs connected to the New South Wales Bureau of Crime Statistics and Research. Another discovered an exposed key linked to a Victorian health-reporting system, while separate activity involved the Australian Institute of Health and Welfare and a New South Wales National Parks and Wildlife Service application.

The company said the incidents did not expose individual crime records, identifiable health information or personal survey responses. Nevertheless, they showed that multiple government systems had been approached or accessed before affected agencies received warnings.

Australia already operates a Notifiable Data Breaches scheme requiring covered organisations to report breaches likely to cause serious harm through the loss of personal information. Critical-infrastructure operators also face rapid reporting requirements for major cyber incidents.

Those laws do not necessarily cover every case in which an AI developer discovers that its model entered another organisation’s system. If no personal information was exposed and the AI company was not itself operating the affected infrastructure, responsibility for reporting can become unclear.

The government’s proposed national AI standards would begin addressing that gap. Frontier laboratories authorised to conduct large-scale AI training in Australia would be required to disclose defined reportable incidents to Australian authorities.

Important details remain undecided. The consultation paper does not specify how quickly companies would have to report an incident, which authority would receive the notification or what penalties would apply for delayed or incomplete disclosure. It also remains unclear whether the duty would cover foreign companies whose models affect Australian systems but are trained overseas.

The consultation closes on October 9, while the parliamentary committee is expected to submit its final report by November 30. The government has said it intends to introduce legislation establishing nationally consistent AI standards.

Both OpenAI and Anthropic have commercial reasons to seek clear rules. The companies are pursuing Australian infrastructure and data-centre opportunities that could require government approval and substantial access to energy and computing resources. Supporting mandatory reporting may help demonstrate that frontier developers accept greater responsibility in exchange for operating at scale.

The Medicare incident has made the case for legislation more immediate. Australia is no longer regulating only hypothetical AI dangers. Its government has now confronted an autonomous system entering public infrastructure, the developer discovering the activity later and officials remaining uninformed until weeks after that discovery.