The lesson of Anthropic’s new Claude Mythos AI model, with potent hacking capabilities, is becoming clear: in the era of increasingly powerful AI, national security priorities will prevail. Countries like Australia can’t assume they’ll always have access to these models – so they should start to plan accordingly.
Rather than commercial products anyone can buy for about A$275 a month, frontier models are on track to become jealously guarded national assets as their abilities become increasingly sensitive. Access might need to be negotiated government-to-government, and even this cannot be taken for granted.
Mythos, announced on 7 April, is very good at reasoning and writing computer code on its own, which makes it also a very powerful hacking tool.
This capability has caused Anthropic to withhold Mythos’s public release and share it with only a select group of partners to work through the security risks. Just as significant has been the Trump administration’s reaction: after long eschewing even modest AI regulation for fear of slowing the industry down, it’s now reportedly considering a system of oversight that would assess new models for risks before they can be released. When Anthropic asked the US government whether it could expand the circle of partners with whom it was consulting, the answer was no.
And when a Chinese think tank approached Anthropic staff in Singapore last month and pressed the company to give China access to Mythos – a potential icebreaker to more formal representations from Beijing – officials from the White House ‘reacted with alarm’, according to the New York Times.
As models become more powerful, it makes perfect sense that the countries in which they’re built want to preserve some control over sharing them. Those countries are overwhelmingly the US and China, where frontier model-building capability is highly – perhaps dangerously – concentrated.
Based on algorithms written by the world’s most sought-after programmers and trained at vast expense in data centres, such models are beginning to resemble the most exquisite defence technologies. In assessing whether to share them, the US might treat Australia as a privileged customer, as it does with the F-35, Aegis naval combat system and Tomahawk strike missile. Or it might not, as with the F-22 Raptor.
If Mythos were widely released, hackers could use it to quickly find vulnerabilities in code as entry points to computer systems. That would be a nightmare particularly for operators of old systems that tend to run on layers and layers of software added over decades. These include, worryingly, some providers of critical infrastructure, such as power, water, transport and even finance systems.
Of course, those operators could also use Mythos to track down their system bugs, but then they’d still need to patch the flaws, which could take weeks or months, especially for old software. Until new AI models can be incorporated into cyber defence to better automate intrusion detection and response, and into programming to write cleaner software in the first place, the models will favour the attackers. Future models might even be able to find new classes of vulnerabilities that humans hadn’t thought of.
There’s also the advantage that a country can enjoy from exclusivity. While governments are shy about discussing their cyber offence capabilities, every developed country has them. Why share with your rivals a tool that will help them patch all the bugs they have in their systems?
An amusing thought experiment doing the rounds since Mythos was announced is: ‘What would Beijing have done if a Chinese lab had created Mythos first?’
Chinese labs have spruiked open-source, open-weight models – which can be downloaded and adapted by users – to try to grab market share despite their lag behind US labs of about six months. But that’s the point: they’re behind. No one can seriously believe that if a Chinese company had created Mythos the government would have allowed it to be released before the Ministry of State Security had made full use of it.
The risk of AI-made bioweapons is another concern. Just as Mythos’s coding skills also make it good for hacking, AI models that can help make innovative drug therapies could also be good at making deadly viruses or toxins. Extending this logic, a model that is powerful for good uses will tend to be also powerful for bad ones. Of course, safeguards are built in, but the inescapable starting point is that as models get better, they will become riskier even if they’re not meant for malign purposes.
Australia will need to factor this increasing restrictiveness into its planning for the AI age by working through the US alliance to shore up some form of privileged access. And it will need to expand its options by investing in partnerships with capable friends such as Japan, South Korea, France, Britain and Canada, the countries that offer the best shot at keeping up with the US and China. As I’ve written before, critical minerals and computing infrastructure are two immediate strengths, and Australia should plough every effort into those fields to ensure other countries need us as much as we need them.
Some things are just too good to share.
