Department secretaries say open-weight AI can be a secure, cost-effective option for routine government work on government infrastructure

The committee brings together departmental secretaries on digital, data and cyber questions, and its own performance report goes to the Secretaries Board. Its communiques are short. This one gives a single two-sentence item to AI models, under the heading Sovereign AI capability: right-sizing the Commonwealth’s AI model portfolio to capture value and reduce dependence.

Members noted that open weight models can provide a secure, cost-effective option for routine government workloads when hosted on government-controlled infrastructure, while frontier models remain important for more specialised use cases.

Secretaries Digital and Data Committee, communique of 24 September 2026

The sentence before it gives the reason: the committee discussed strengthening the Commonwealth’s AI resilience, noting risks associated with reliance on a small number of overseas AI providers. The communique names no model, no workload and no cost. Under adoption, it also records that evidence of benefits remains emerging and uneven across the APS, and that Finance is working with the Digital Transformation Agency (DTA) and agencies on a shared way to measure them.

How much AI the public service already uses

At the Joint Select Committee on Artificial Intelligence on 30 October, the DTA and Finance gave a joint opening statement with current figures. Finance runs GovAI, the shared platform, which they described as a multimodel, multiprovider design that reduces reliance on any single supplier or technology.

The hosting figure matters for the secretaries’ idea. An open-weight model on government-controlled infrastructure is something an agency, or GovAI, has to run itself, and 13 agencies host AI solutions through GovAI today.

What the security rules say now

At the same hearing, Home Affairs officers were asked whether they were doing any work comparing the safety and security of open-weight and closed-weight models. Timothy Neal, an acting first assistant secretary, said the department and the Australian Signals Directorate (ASD) had advised on open-weight deployment since the 2025 direction on DeepSeek, through a policy explanatory note issued with it. (We have written about that direction and the deny list behind it.) Greta Doherty, a first assistant secretary, said the wider question, for government, critical infrastructure and the economy, is open: That is work that we are currently working through.

Asked what the rules are in the meantime, Mr Neal said he would take it on notice, then sketched them. Agencies need the capability to run the models, take a risk-based approach, and follow ASD guidance on deployment. Where security classified information is involved, Agencies must use those certified environments when deploying open-weight models, and they need to come to the Department of Home Affairs and the Australian Signals Directorate to demonstrate that they have that capability and capacity. He said the capability of the agency is the first and most important component.

Summarised from the communique and the Proof Committee Hansard of 30 October 2026, pages 38, 41, 43 and 62.

The AI Safety Institute’s caution

Later that day Dr Kate Conroy, general manager of the AI Safety Institute inside the Department of Industry, Science and Resources, was asked why Australia could not simply use open-weight models with Australian data. She said the institute had been following them closely and There are a lot of red flags that are coming out now in the evidence with regard to significant security risks from using open-weight models. She named easy fine-tuning to make a model behave in ways its user would not want, and manipulated models being made available for download. She also said There are going to be circumstances where open-weight models will be totally fine. Her conclusion on secure and air-gapped use: There are workarounds here, but they are not as simple as a replacement.

The transcript is a proof, which witnesses can still correct.

Our view

The direction is reasonable. A public service that sends every routine task to two or three overseas providers has a concentration risk, and the secretaries have named it. An open-weight model run inside government infrastructure keeps the data and the off switch at home, and GovAI was built to offer more than one provider.

But the communique is two sentences, and the evidence beside it is thinner than its confidence. The security agencies attach conditions the communique leaves out: Home Affairs makes an agency’s own capability the first test, is still settling its broader posture, and took its current rules on notice, and the AI Safety Institute’s general manager led with risk. “Cost-effective” arrives with no figure, in the same communique that says the evidence of AI benefits across the public service is still emerging and uneven. The test will be whether the secretaries’ December meeting, or Finance’s GovAI work, publishes what counts as a routine workload, which models qualify, and what they cost against the overseas services they would replace.

Two dates to watch. Home Affairs’ answer on the current rules is a question on notice to the joint committee, whose answers are published on its website. The committee must present its final report by 30 November.