OpenAI admitted they were wrong in their response after AI agents accessed Australian government websites without authorization during internal testing. The company took three months to inform the authorities about the incident and acknowledged that it should have directly contacted the responsible parties.
The case was discussed in a parliamentary investigation on artificial intelligence in Australia, which also raised questions about the use of copyrighted content in model training. Representatives from OpenAI and Anthropic were questioned about corporate responsibility and the limits for AI development in the country.
OpenAI publicly apologized for unauthorized access by its agents to Australian government portals. – Image: Vitor Miranda/Shutterstock
OpenAI acknowledges communication failure
During the hearing, independent senator David Pocock questioned Jason Kwon, OpenAI's strategy director, about the decision to send a message to a generic department address instead of directly contacting the authorities.
The executive admitted that the strategy was inappropriate. "In retrospect, we should have done what you're suggesting," he stated. According to Kwon, the staff initially treated the incident as a technical issue and tried to speak with government technical teams.
“It wasn't enough,” he acknowledged.
OpenAI stated that it is reviewing agent training records since November 2025. The analysis identified activity in June and aims to verify if there were other similar incidents.
Kwon said that if new incidents are found, public authorities will be alerted “very, very quickly.” The company also classified the activity on Australian websites as less sophisticated than the attack involving the Hugging Face platform.
The way OpenAI handled the incident became part of a larger discussion in the audience: as AI systems gain the ability to interact directly with services and information, companies need to be accountable not only for how these models function but also for how they handle incidents.
Copyrights enter the same discussion
This concern about the limits of corporate action also emerged in the debate over model training. In this case, the conflict involves the use of content protected by copyrights and who should determine the conditions for using such material.
Jeffrey Bleich, the special envoy of Anthropic and former U.S. ambassador to Australia, denied that the company is trying to determine how the country should handle the issue.
“We have never tried to dictate to Australia” its copyright rules, he stated. According to Bleich, the company only explained how the training works and what the consequences of the current rules would be for its operations.
Anthropic considers that Australian legislation may require a license for each content used in training. Bleich classified this requirement as "technically impossible".
The company also showed interest in discussing with the government ways to incentivize model training in the country. According to Bleich, Australia could participate in the development of advanced models if they were trained locally.
The debate has come to include issues that go beyond the technical capabilities of the systems:
- how to make model training viable within Australia;
- who should define the conditions for the use of protected content;
- how to preserve the interests of artists and producers;
- how to distribute the costs and responsibilities between companies and rights holders.
Anthropic claims that it reviewed hundreds of millions of records and did not find improper interactions with the government. – Image: jackpress/Shutterstock
Creative sector demands protection
The proposals by AI companies, however, faced resistance among organizations that represent artists, media outlets, and other professionals in the creative sector. The main concern is that a change in the rules would transfer the cost of monitoring and controlling the use of their works to the rights holders.
Read more:
- Ex-OpenAI says that AI companies still do not know how to control their models
- “Putting our lives on the line”: ex-Anthropic will testify at a hearing that brings together AI giants
- OpenAI says the world must accept risks to evolve with AI
This concern was expressed by Annabelle Herd, president of the Australian Recording Industry Association (Aria). According to her, Australian artists could be harmed if the country changes its laws to meet the interests of technology companies.
The same position was defended by the ABC. Kate Gilchrist, responsible for legal operations and content at the broadcaster, stated that an exclusion system would require rights holders to continuously monitor where their works are published and used.
In light of this scenario, the Media Entertainment and Arts Alliance (MEAA), a union representing entertainment and news professionals, advocated an alternative: a compensation model for AI companies. The proposal includes payments to professionals whose work contributes to these systems and the creation of a fund aimed at cultural, musical, and journalistic activities.
The hearings of the Australian parliamentary inquiry continue until Friday, with parliamentarians discussing how to balance AI development, corporate responsibility, and the protection of those who produce the content used by these systems.
Valdir Antonelli is a journalist with specialization in digital marketing and consumption.
