In the wake of yesterday’s news that an OpenAI agent hacked Australia’s Medicare system earlier this year, some people have pointed to the incident as evidence that Australia needs to develop its own artificial intelligence models.
For example, Assistant Minister for Science, Technology and the Digital Economy Andrew Charlton said the breach:
provides the justification … for our determination to make sure that we have Australian AI, that we shape technology, we have sovereign capability in Australia, and we aren’t entirely at the mercy of foreign AI companies”.
Deputy Opposition Leader Jane Hume echoed this. Speaking on Sunrise this morning, she said the hack shows why Australia needs to embrace major AI companies “so that we can have frontier models providing sovereign capability to Australia”.
But would having an Australian AI actually help protect us from an AI agent-led hack? And if so, how?
The first known hack of a government system
As part of a training exercise in June, OpenAI tasked an agent with researching public medical data. In pursuit of this goal, the agent breached a firewall to gain access to an old Australian government website that contained private statistical data.
OpenAI became aware of the hack in August during a review of the model’s activity. It only notified the Australian government – via an email to a low-level public mailbox – on September 10.
Prime Minister Anthony Albanese has announced a rapid taskforce to provide an “urgent and immediate review” of the incident, and whether changes to the law are needed to prevent another incident like this.
This is the first known occasion of an AI agent breaching a government system. But it follows several other high-profile recent incidents of AI agents autonomously hacking a third-party system.
Would an ‘Australian AI’ help?
In the context of AI, sovereign capability is where countries develop and maintain the technology themselves, rather than relying on models that are built and based elsewhere. Complete AI sovereignty means a country has independent control over not just an AI model, but also the data and physical infrastructure behind it.
In theory, building “Australian AI” is a good idea. It mitigates the risk of other countries using their AI models as political leverage, and would give us more autonomy.
But there are significant practical challenges when it comes to developing sovereign AI. The first is the financial cost.
If Australia tried to develop and train a frontier AI model that could perform as well as commercial equivalents such OpenAI’s ChatGPT or Anthropic’s Claude, it would cost upwards of a billion US dollars.
It would also require us to build our own data centres to train an Australian AI model, rather than using the existing centres abroad that are owned by international corporations such as Amazon Web Services or Microsoft Azure. This would consume massive amounts of power and water. Epoch AI, a non-profit investigating the future of AI, estimates that by 2030, training individual frontier AI models could draw between 4 and 16 gigawatts of power – enough to power millions of homes.
Let’s say the Australian government went against expert advice and chose this approach. Would it better protect us from incidents such as the OpenAI Medicare hack?
Unfortunately, the likely answer is no.
Many other countries, such as South Korea, Switzerland and France, have tried to compete with leading AI companies by building their own domestic AI capability. But none of the AI models these countries have built are anywhere near as advanced as the leading Chinese and US models.
It is those highly advanced models, which are continuing to improve, that pose the greatest cybersecurity risk – as the spate of recent AI agent-led hacks demonstrates. And an Australian-built AI model that doesn’t compete on performance would do little to mitigate that risk. It would likely miss cyber vulnerabilities that a more advanced model could still identify and exploit.
So, what do we do?
The only way to guarantee Australia isn’t the victim of another AI cybersecurity incident is to completely cut ourselves off from the global internet. Obviously, that isn’t a real option.
It’s important Australia keeps pushing for stronger global regulation and rules governing AI systems, as Albanese has been doing in New York this week. AI isn’t the first dangerous technology humans have created, and the international approach to governing aviation and nuclear weapons, while imperfect, provides a suitable model for how AI could be better regulated to mitigate the risk.
These rules should be focused on improving transparency of AI models, for example through mandatory independent testing which could recommend additional safeguards.
It’s also crucial that Australia holds AI companies accountable for cybersecurity breaches, as other jurisdictions around the world have done. The government has already flagged it may press criminal charges against OpenAI over the Medicare hack.
Such a move would send a strong message to AI companies that it’s not okay to move fast and break things in the pursuit of bigger and better models.
-
-
-
-
