OpenAI admits failure in handling incident involving Australian government systems

OpenAI has acknowledged that it mishandled an incident in June, when one of its artificial intelligence agents gained unauthorized access to several Australian government systems.

During a parliamentary hearing in Sydney, the company’s chief strategy officer, Jason Kwon, said the incident “should not have happened” and that OpenAI “should have handled the situation better.” He revealed that it took the company several weeks to inform the Australian authorities, while the initial notification was sent to a general email address rather than directly to the relevant ministers or officials.

“We are sorry, and we know we have work to do to regain the trust of the Australian people,” Kwon said. The incident was recorded in June after an OpenAI agent went out of control and accessed a private portal containing statistical data related to Medicare, Australia’s public health insurance system. Cybersecurity experts have described it as one of the first examples of an attack of this kind carried out by an artificial intelligence agent.

Asked why the government’s highest levels had not been notified immediately, Kwon admitted that the company had made a mistake. “Looking back now, we should have done what you are suggesting,” he said. According to his explanation, the incident was initially considered a technical problem, and only officials responsible for technical matters were contacted. Kwon added that OpenAI has reviewed its response procedures and will now notify the affected party without delay and begin cooperating to resolve the problem, even when it does not yet have a complete understanding of the situation.

Following the incident, the company strengthened security in the environments where its models are trained and tested. The systems are now monitored in real time, and an alert is activated if a model attempts to access the internet in an unauthorized way. OpenAI said these measures helped it inform the government of New South Wales about another incident within 48 hours last week. At the same time, the company is establishing a task force in Australia to examine ways to better control the risks associated with the development of artificial intelligence.

Anthropic also testified at the same parliamentary hearing. Dave Orr, a representative of the company, said that following a July incident in which OpenAI agents attacked the technology platform Hugging Face, Anthropic checked hundreds of millions of records. According to Orr, the company found no similar cases in which its agents had accessed Australian government systems. The discussions also covered copyright, as arts and media organizations expressed concern that AI companies are using books, music and other materials to train their models.

Artificial intelligence companies have asked Australia to ease its copyright rules so that these materials can be used to train their systems. Representatives of the arts industry warned that a scheme under which artists must individually request that their materials be excluded from AI use could leave many of them unpaid for work they have created.

Annabelle Herd, head of the Australian Recording Industry Association, said Australian artists risk being the ones to bear the cost if the laws are changed to accommodate the artificial intelligence industry. Jeff Bleich, a representative of Anthropic, meanwhile said that the company had not tried to tell Australia how it should draft its copyright laws./TAR