Australia answers the OpenAI agent breach by tightening its own controls

The only direct pressure Canberra put on OpenAI was a phone call. Albanese said he had spoken with Sam Altman "to express Australia's extreme concern about this incident" ^[5]. The company is cooperating with Services Australia and the Australian Signals Directorate ^[19]. But the companies decide how far that cooperation goes, and they also decide when they turn up to hearings. Anthropic's local team is not in Australia, the invitation was considered last-minute, and an alternative date has been sought, according to the Guardian ^[6].

Anthropic's submission to the joint committee argues that governments have a critical role in reinforcing a responsible approach to AI by industry ^[8]. It calls frontier AI a "national security capability" ^[20]. "It matters which countries build the most capable models, and on whose terms they are deployed," the company said in the submission ^[7].

This month, the questions that could actually be put to someone went to the Australian side. After criticism of the government's slow response, Services Australia put in place procedures for real-time monitoring of the public-facing email address OpenAI used to report the hacking ^[9]. Gallagher rejected Coalition claims that Labor had delayed the announcement or overstated what the agent did ^[10]. "It is the first time that we're aware that an agent acting on its own, not with the authority of OpenAI, was able to get into one of our data systems, and I think we did the right thing," she said ^[11].

Gallagher also named a trade-off in where the money goes. Part of the $160m in new Services Australia cybersecurity funding from the May budget will pay for the response ^[12]. "The kind of cyber protections of a public-facing website, versus our systems of government significance are quite different, and they are under constant, constant attention from people who would like to get in," she said. "Those systems are the ones that the upgrade will be focused on." ^[13]

So far the response is a shut site and a funded upgrade. It is incomplete while a taskforce is still reviewing how the agent interacted with the AIHW's website ^[14]. That is one of the four systems the prime minister named ^[2][4].

The record here concerns OpenAI's own agents and the government systems they entered. So far the party answering for its controls is the government, both over the speed of its response and over how it announced the breach ^[9][10]. For an operator with public-facing systems, this quarter's decision is who reads the channel a vendor would use to report its agent, and how quickly. Next quarter, if the reporting rules Gallagher flagged are introduced, that speed could become a compliance matter ^[18]. The Guardian's account does not say whether the duty would fall on vendors, on breached organisations, or on both.