- Scope
- Australian business access and continuity; policy position stated 30/09/2026
- Article assessment confidence
- High on the stated objective; implementation feasibility remains open
My position
This article explores what Australian control of AI could usefully achieve. Inference, using a trained model to produce an output, is one part of that question. Control over inference matters, but I would not assess it in isolation from the capabilities and infrastructure it depends on, or how those might change over time. Training models in Australia can have benefits, but it does not by itself settle who controls their subsequent use. A local building and local data storage do not settle that question either.
One test I would apply is whether an Australian operator can legally and practically keep running an agreed model if an overseas partner withdraws access. Different arrangements may meet that objective. I have not prescribed a single ownership structure or technical design.
Why continuity deserves attention
Continuity matters when AI becomes part of the infrastructure a business or country relies on. Cybersecurity makes the stakes clear. AI-enabled attacks can operate at a speed and scale that human teams cannot match unaided. Defenders need tools that can keep pace, including AI. If a business or country relies on AI for its defence, losing access is a security risk. Who controls that access becomes a practical security question.
On 12/06/2026, Anthropic suspended Fable 5 and Mythos 5 for all customers following a US directive restricting foreign-national access. Other models remained available. Anthropic subsequently announced that the controls had been lifted and access restored, with an update on 01/07/2026. The interruption was temporary, but the dependence was real.
I want Australia to have an option it can rely on for essential work, including cyber defence, even when an overseas supplier or government changes the terms of access.
What this does not require
I support partnerships with overseas developers, including US and Chinese providers, subject to the security and control required by the particular arrangement. Businesses should be able to select a service that suits their work. Some may accept a premium for Australian control and continuity; others may choose differently.
Open weights do not automatically deliver sovereignty. Rights to operate, required infrastructure, dependencies and practical capability still need examination. Conversely, an overseas partner should not be excluded simply because it is overseas. The proposal should be assessed on what it actually delivers.
What would make the option useful
Capability has to be judged against the job. A model that is useful for drafting business emails may still be inadequate for cyber defence. In cybersecurity, the test moves as attackers’ capabilities improve. Keeping an older model running is valuable only to the extent that it can still do the work we need. Australia needs a way to maintain and improve the defensive capabilities it relies on, alongside the people and systems needed to use them safely.
Training and implementation support belong alongside infrastructure. Nominal access is a weak outcome if firms cannot put it to work.
I support substantial public investment alongside private investment to secure capability Australia deliberately wants to control. I do not support general subsidies for AI consumption. Users should remain responsible for the services they use.
Continuity and the moving frontier
Keeping an agreed model running can preserve useful capability. It does not by itself secure access to future improvements. For cyber defence, that distinction matters because the threat keeps changing. I want Australian control to include a credible path to maintaining the capability we depend on, rather than just keeping a fixed model available. That requires examining licences, dependencies, updates and practical capability. I explore the risk of compounding research advantages in Can Australia benefit from AI without controlling the frontier?.
Why this edition changed
Reframed the opening as an exploration, made continuity one test, and added the owner-approved cyber-defence and evolving-capability argument. Removed current-argument appeals to historical submissions and the two closing sections at the owner’s request. Historical records remain unchanged.
Before: Every Australian business should have a practical option to use capable models served here under Australian control. The test is credible legal and operational continuity, while retaining international partnerships and the freedom to use overseas services.
Now: Every Australian business should have a practical option to use capable models served here under Australian control. The test is credible legal and operational continuity, while retaining international partnerships and the freedom to use overseas services.
Confidence: High on the stated objective; implementation feasibility remains open
Practical effect: Readers can assess continued access against the work a model must do, distinguish current views from dated submissions, and follow the new sources and preserved editions. No new confidence level or separate view-register change is made.
Edition history
v3 · 19/10/2026 · Edition source
Sources and limits
Coordinated contracting arrangements principles
Addresses operational authority, foreign-law exposure, removal of services, portability and exit. Useful questions for continuity planning. It does not establish that an Australian contract overrides foreign law or automatically apply every Commonwealth procurement rule to SMEs.
Statement on the US government directive to suspend access to Fable 5 and Mythos 5
Provider reports temporarily suspending two models for all customers in response to a foreign-national access directive. Other Anthropic models remained available. Read alongside the restoration announcement; it was not a permanent or industry-wide outage.
Redeploying Claude Fable 5
Announces lifting of the export controls and restoration of access, with update dated 01/07/2026. A provider account of a particular event, not a prediction that Australian access to all AI will be withdrawn.
Disrupting the first reported AI-orchestrated cyber espionage campaign
A provider account of one campaign, rather than an independent or universal measure of AI attacks. It reports speed and parallel activity beyond unaided human operation while retaining human direction and noting model errors. It does not establish fully autonomous attacks in all settings.
Impact of AI on cyber threat from now to 2027
A formal assessment of likely developments through 2027, not a measured Australian effect or a guarantee. It judges keeping pace important for resilience, expects human involvement in advanced attacks, and warns that insecure AI integration creates additional risks. It does not establish that AI alone is the only defence or that domestic control ensures effectiveness.
One does not simply defend agentically
Explains the organisational constraints and risks of defensive automation. Detection and tightly scoped responses require responsibility, human action and safeguards against harming the organisation. It qualifies the cyber-defence argument; it is not evidence that sovereign AI by itself guarantees effective defence.