Secretaries back open-weight AI for routine government work on sovereign infrastructure
A communique from the Secretaries Digital and Data Committee says open-weight models can be a secure, cost-effective choice for routine workloads if government hosts them. Home Affairs and the AI Safety Institute told a parliamentary committee the security picture is still unsettled.
Department secretaries have told government that open-weight AI models could handle routine workloads securely and cheaply, as long as they run on infrastructure the government controls. The Secretaries Digital and Data Committee met on 24 September 2026 and released its communique on 6 November. Frontier models, it said, stay important for more specialised uses.
The stated reason is the risk of depending on a small number of overseas AI providers. The communique does not name a model, a workload or a cost. It also notes that evidence of benefits from AI across the Australian Public Service is emerging and uneven. Finance is developing a common method for measuring those benefits with the Digital Transformation Agency (DTA) and agencies.
The DTA and Finance gave a joint opening statement to the Joint Select Committee on Artificial Intelligence on 30 October, with figures as at October 2026. They put the number of agencies and organisations engaged with GovAI Chat at 74, and staff using its chat function at more than 12,000. Thirteen agencies host AI solutions through GovAI, 133 chief AI officer appointments have been notified to Finance, and more than 100,000 staff have finished AI in government fundamentals training. Finance runs GovAI, which the two agencies described as a multimodel, multiprovider design that lessens reliance on any single supplier or technology. The hosting figure is relevant because an open-weight model on government infrastructure must be operated by an agency or by GovAI.
Asked at the same hearing about safety and security comparisons of open-weight and closed-weight models, Home Affairs acting first assistant secretary Timothy Neal said his department and the Australian Signals Directorate (ASD) had advised on open-weight deployment since the 2025 direction on DeepSeek. Asked for the current rules, Neal said he would take the question on notice, then sketched them. Agencies need the ability to run the models, must take a risk-based approach and must follow ASD deployment guidance. Where security classified information is involved, agencies must use certified environments and show Home Affairs and ASD they have the capability and capacity. Neal said an agency's own capability is the first and most important component. First assistant secretary Greta Doherty said the wider question for government, critical infrastructure and the economy remains open and is work still under way.
Later that day, Dr Kate Conroy, general manager of the AI Safety Institute within the Department of Industry, Science and Resources, said the institute had been following open-weight models closely. She said emerging evidence raised many red flags about significant security risks, naming easy fine-tuning that can make a model behave in unwanted ways and manipulated models offered for download. Conroy also said open-weight models will be fine in some circumstances, and that workarounds exist for secure and air-gapped use but are not as simple as a replacement. The Hansard transcript is a proof that witnesses can still correct.
Home Affairs' answer on the current rules will come as a question on notice to the joint committee, with answers published on its website. The committee must present its final report by 30 November.