National security priorities will prevail: countries like Australia can’t assume they’ll always have access to frontier AI models, so they should start to plan accordingly.
Anthropic's Mythos, its handling, and the US response
Mythos, announced on 7 April, was described by its maker as very good at reasoning and writing computer code on its own. Anthropic withheld Mythos’s public release and instead shared it with only a select group of partners while the company worked through security risks. The firm’s approach contrasts with lower-tier commercial products that can be bought for about A$275 a month.
The United States government has stepped into the picture. The Trump administration is reportedly considering a system of oversight that would assess new models for risks before release, and when Anthropic asked the US government whether it could expand the circle of partners with whom it was consulting, “the answer was no.” The New York Times also reported that when a Chinese think tank pressed Anthropic staff in Singapore to give China access to Mythos, White House officials “reacted with alarm.”
Why Mythos tilts the cyber balance toward attackers
Mythos’s ability to write and reason about code makes it a potent hacking tool: if widely released, hackers could use it to quickly find vulnerabilities in code as entry points to computer systems. That is especially alarming for operators of old systems that run on decades of layered software additions — examples cited include some providers of power, water, transport and finance systems.
Operators could also use Mythos to track down bugs, but patching flaws in legacy systems could still take weeks or months. Until new AI models are incorporated into cyber defence to better automate intrusion detection and response, and into programming to write cleaner software, the models will favour the attackers. The analysis adds that future models might even be able to find new classes of vulnerabilities that humans hadn’t thought of.
The piece also flags biological risk as part of the same logic: models that can help make innovative drug therapies could, in theory, also help make deadly viruses or toxins. Safeguards exist, the article notes, but the inescapable starting point is that as models get better, they will become riskier even if they aren’t meant for malign purposes.
Frontier models as national assets — control, concentration, and precedent
Frontier model-building capability is concentrated overwhelmingly in the US and China, and based on algorithms written by top programmers and trained at vast expense in data centres, these models are beginning to resemble defence technologies. The piece argues it makes sense for countries where models are built to preserve control over sharing them.
That control could take the form of privileged government-to-government access — the author compares potential treatment to how certain weapons systems are shared: 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. Conversely, Chinese labs have promoted open-source, open‑weight models to capture market share despite lagging US labs by about six months — but the article poses the thought experiment of what Beijing would have done if a Chinese lab had created Mythos first, suggesting security services would have sought exclusive use.
What this means for Australia, the US alliance, and critical-infrastructure operators
- Australian policymakers and planners: They should factor increasing restrictiveness into planning by working through the US alliance to shore up some form of privileged access, and by investing in partnerships with Japan, South Korea, France, Britain and Canada as alternate bilateral tracks.
- Operators of critical infrastructure and security teams: They face a near-term risk that attackers using advanced models will find and weaponize vulnerabilities in legacy systems; while defenders can also use advanced models, patch timelines for old software mean exposure remains significant.
- Allied technology partners (Japan, South Korea, France, Britain, Canada): These countries offer “the best shot at keeping up with the US and China,” so partnership and capability-sharing with them is presented as a practical route to expand Australia’s options beyond reliance on any single supplier.
Where Australia can act: critical minerals, computing infrastructure, and negotiation
The article identifies two immediate strengths Australia can and should exploit: critical minerals and computing infrastructure. It argues Canberra should “plough every effort into those fields to ensure other countries need us as much as we need them.” That approach is offered as a practical complement to diplomatic work through the US alliance — a way to create bargaining leverage if frontier AI models become jealously guarded national assets.
Australia’s core choice, the piece concludes, is whether to treat access to frontier models as something to negotiate for, or to try to replace through sovereign capability and alliances. The stakes are framed concretely: without planning, Australia cannot assume it will always have access to the most powerful AI tools and the defensive benefits they could bring.
