\n\t\t\tRelated Article Block Placeholder<\/p>\n
\t\t\t\tArticle ID: 338862\n\t\t<\/p>\n
\t\t\t\t\t\t\t Within hours, those models disappeared for users outside US jurisdiction \u2013 only days after Mythos was first made available.<\/a><\/p>\n And it wasn\u2019t long before it happened again.<\/p>\n OpenAI announced GPT\u20115.6<\/a>, a new model family led by its flagship Sol, designed to handle more complex coding and cybersecurity tasks.\u00a0<\/p>\n \tSmarter business news. Straight to your inbox.<\/p>\n \n For startup founders, small businesses and leaders. Build sharper instincts and better strategy by learning from Australia\u2019s smartest business minds. Sign up for free. <\/p>\n By continuing, you agree to our Terms & Conditions<\/a> and Privacy Policy<\/a>.<\/p>\n But instead of a broad launch, OpenAI said access would start with about 20 \u201ctrusted partners\u201d, pre\u2011approved by the Trump administration, after the White House asked it to limit the release.\u00a0<\/p>\n OpenAI said in its announcement it doesn\u2019t want this sort of government access process to become the default but is accepting it as a short-term step.<\/p>\n At the same time, Anthropic has been warning Washington about what others were doing with its models.\u00a0<\/p>\n In a letter to US senators<\/a>, it accused operators linked to Alibaba of running the \u201clargest known distillation attack\u201d on Claude by using tens of thousands of fake accounts and millions of queries to train their own systems.\u00a0<\/p>\n Anthropic framed this as \u201cillicit\u201d capability theft that turned US AI investment into a subsidy for a geopolitical rival and asked Congress to crack down.<\/p>\n In both the Anthropic and OpenAI cases, national security decisions in Washington are now shaping which tools Australian organisations can use and when.\u00a0<\/p>\n What has been made abundantly clear is that all the talk about AI being borderless and the great equaliser is bogus. Access and control of these models come firmly under US jurisdiction, no matter where in the world they\u2019re being used.<\/p>\n They\u2019re also not mere neutral cloud tools. Frontier models are being treated as strategic assets whose release, access, and even imitation are questions of state interest.<\/p>\n For Simon Kriss, CEO of Sovereign Australia AI<\/a>, that should end any illusions about how much control Australia really has.\u00a0<\/p>\n \u201cI see this as Australia\u2019s wake-up call. So long as we rent our future from the US or China, we are at their whim,\u201d he told SmartCompany.\u00a0<\/p>\n He also points out the practical cost. Governments and businesses may need \u201cmodern-day switchboards\u201d that can detect when a model disappears and instantly flip to another.<\/p>\n \n\t\t\tRelated Article Block Placeholder<\/p>\n \t\t\t\tArticle ID: 338762\n\t\t<\/p>\n \t\t\t\t\t\t\t Kriss says this is \u201cnot cheap, easy or quick\u201d and may change the economics of rolling out agentic AI processes at scale.\u00a0<\/p>\n Across the G7, these US orders have already been read as a warning shot<\/a>. Governments in Europe and Canada have pointed to the Anthropic incident as justification for accelerating their own sovereign AI and compute plans.<\/p>\n It\u2019s worth noting that as this column was being finalised, the US Commerce Department lifted the export controls<\/a> on Anthropic, less than three weeks after imposing them.<\/p>\n This follows a partial rollback over the weekend where, similar to OpenAI, the government allowed Mythos 5 to be reinstated for a handful of \u201ctrusted\u201d US organisations.<\/p>\n Now Anthropic will start restoring access more broadly, after agreeing to tighter security protocols and early notification of any \u201cmalicious activity\u201d.<\/p>\n However, the fact that Anthropic\u2019s export controls were lifted within weeks doesn\u2019t make the situation less serious. Both shutdowns and restorations of frontier AI access happen on someone else\u2019s timetable.<\/p>\n AI sovereignty versus agency<\/p>\n In Canberra, the term that keeps cropping up is \u201csovereign AI\u201d.\u00a0<\/p>\n There\u2019s been an uptick in ministers and shadow ministers alike working the phrase into speeches about national capability and the digital economy in recent years.\u00a0<\/p>\n Just last week<\/a>, Industry Minister Tim Ayres warned Australia cannot afford to be \u201ca cork bobbing on the ocean of other people\u2019s technology\u201d and should secure as much of the technology stack as possible.<\/p>\n Opposition figures like Andrew Hastie have started framing AI as a strategic contest reminiscent of the Cold War arms race<\/a>. Hastie warned under-investment risks constraining Australia\u2019s sovereignty and independence.<\/p>\n Researchers are trying to give that rhetoric some tangibility.\u00a0<\/p>\n The Automated Decision-Making and Society centre breaks sovereignty down<\/a> into four elements: data, compute, model and policy.<\/p>\n Across those metrics, Australia has seen patches of progress. There are on\u2011shore cloud and hosting providers, government-focused platforms, and a growing pipeline of AI-friendly data centres.\u00a0<\/p>\n What the Anthropic shutdown exposed is that at the frontier, Australia still lacks model and policy sovereignty when it comes to the LLMs that are being sold to us as state of the art.<\/p>\n Decisions about access sit with another government.<\/p>\n The Tech Policy Design Institute (TPDi) refers to this as \u2018AI agency\u2019. This is a country\u2019s ability to control outcomes, protect its interests and capture value from AI across infrastructure, data, models, skills, innovation and governance.\u00a0<\/p>\n The TPDi\u2019s recent report<\/a> on the issue assesses 103 capability areas across six layers, finds that Australia\u2019s agency is constrained by gaps in frontier models and domestic compute, even as it scores relatively better on skills, innovation and adoption.<\/p>\n But Kriss is sceptical of attempts to soften the language.<\/p>\n \u201cMany people in Australia are using the word \u2018agency\u2019 to get around an inconvenient conversation on true sovereignty,\u201d he said.<\/p>\n \u201cHow can you have any form of agency if someone in another country can just turn off the tap?\u201d<\/p>\n Inside Sovereign Australia AI, he talks about five levels of sovereignty \u2013 infrastructure, models, agentic tools, applications and governance.<\/p>\n But the underlying idea is relatively straightforward \u2013 GPUs and servers owned and operated here, models built and run here, and rules set here.<\/p>\n \n\t\t\tRelated Article Block Placeholder<\/p>\n \t\t\t\tArticle ID: 322958\n\t\t<\/p>\n \t\t\t\t\t\t\t \u201cThe only answer is our own AI on our own servers that is owned and operated by Australian entities only,\u201d Kriss said.\u00a0<\/p>\n Some builders, however, have lost patience with the glossary of terms.<\/p>\n Maincode\u2019s view on dependence<\/p>\n Maincode,<\/a> which has created the Matilda model as well as a Melbourne-based \u2018AI factory\u2019, stopped using the term \u2018sovereign AI\u2019 last year.<\/p>\n \u201cWe ditched the term because it doesn\u2019t mean anything. It\u2019s not the language of builders. You can\u2019t test it; it holds no value, no meaning,\u201d Maincode CEO Dave Lemphers told SmartCompany.\u00a0<\/p>\n Instead, Maincode talks about \u201cAustralian-made AI\u201d that has an end\u2011to\u2011end stack built and run locally.<\/p>\n Lemphers says the Anthropic order did not materially affect Maincode because the company is not relying on US-hosted frontier models in the first place. Instead, it runs on its own models that he claims outperform Claude.<\/p>\n \u201cWe use our own models and consume our own tokens, from chat through Matilda to our Matilda-code coding agent,\u201d Lemphers said<\/p>\n \u201cI\u2019ll admit it does concern me, though, that there are still Australian organisations, both government and enterprise, comfortable with a foreign political entity like the White House dictating what AI they get to use.<\/p>\n \u201cFor us at Maincode, it\u2019s liberating to not have to worry about access to state-of-the-art AI being dictated by someone else.\u201d<\/p>\n The company has detailed publicly how it wrote its own inference engine in Rust for AMD Instinct GPUs, bypassing the standard vendor stack. It also argues owning an on\u2011premise \u201cAI token factory\u201d makes tokens far cheaper than buying them through public APIs.<\/p>\n \u201cThe reality is, we\u2019re doing it, and doing it well. People talk about it in terms of \u2018what if\u2019. We\u2019re so past that. We\u2019re actually doing it, and I think the conversation is lagging behind the reality.\u201d<\/p>\n Lemphers doesn\u2019t mince words when it comes to Australia\u2019s reliance on offshore models.<\/p>\n \u201cThe idea that using foreign-controlled AI is completely fine goes far beyond commercial considerations. It is the single biggest standalone risk to Australia\u2019s national security posture and our economic ability to compete globally,\u201d he said.<\/p>\n \u201cI just don\u2019t think those in power actually get it, to be honest.\u201d<\/p>\n What this means for Australian businesses<\/p>\n This might sound like a story for policymakers. It\u2019s not.<\/p>\n This sits firmly in the territory of how Australian businesses buy and use AI.<\/p>\n If a company has built workflows around a specific US\u2011hosted model, it has assumed that model will be there on roughly the same terms tomorrow. The Anthropic shutdown showed that assumption can be overturned overnight.<\/p>\n If an AI vendor is selling itself as Australian or \u201con\u2011soil\u201d but cannot say what happens when US or other export rules change, that\u2019s a red flag.<\/p>\n \n\t\t\tRelated Article Block Placeholder<\/p>\n \t\t\t\tArticle ID: 339545\n\t\t<\/p>\n
\n\t\t\t\t![\"Why](\"https:\/\/www.newsbeep.com\/au\/wp-content\/uploads\/2026\/07\/20260507112177280044-original-1.jpg\")
\t\t\t<\/a><\/p>\n
\n\t\t\t\t![\"Neural](\"https:\/\/www.newsbeep.com\/au\/wp-content\/uploads\/2026\/07\/Copy-of-SCM-ed-image-1920x1080-16.png\")
\t\t\t<\/a><\/p>\n
\n\t\t\t\t![\"Neural](\"https:\/\/www.newsbeep.com\/au\/wp-content\/uploads\/2026\/07\/Copy-of-SCM-ed-image-1920x1080-22.png\")
\t\t\t<\/a><\/p>\n
![\"We](\"https:\/\/www.newsbeep.com\/au\/wp-content\/uploads\/2026\/07\/Copy-of-SCM-ed-image-1920x1080-26.png\")
\t\t\t<\/a><\/p>\n