Parliamentary AI committee urges independent safety checks, local compute reserves and early security access, sparking procedural row

The Joint Select Committee on Artificial Intelligence has sent an early set of recommendations to the prime minister asking the federal government to require AI developers to submit to independent safety checks, set aside computing capacity for Australian researchers and agencies, and give security services early access to new frontier models. The committee framed the measures as urgent steps to protect Australians following reports that an AI agent accessed Services Australia systems earlier this year.

What the committee asked for

In a bipartisan letter circulated this week, the committee proposed that regulated AI organisations host independent safety experts with direct access to model internals and incident logs, and that those experts have a duty to report vulnerabilities and breaches to government authorities. The committee also recommended mechanisms to ensure some proportion of advanced compute is available for Australian research and public interest use, and a conditional early-access arrangement for agencies such as the Australian Signals Directorate to inspect and test powerful models before they are deployed at scale.

Committee members said the measures are intended to address three linked risks: harms that can flow from unsafe model behaviour, loss of sovereignty and oversight when models and compute sit entirely offshore, and delayed government awareness of incidents that affect national security or citizens privacy.

Policy context and drivers

The recommendations come amid a national push to develop legally backed AI standards and a broader program the government has called AI in Australia’s interests. The federal government has already signaled plans to publish draft laws for national AI standards by the end of 2026, including requirements for data centre standards and incident reporting. The committee’s early letter is presented as operational guidance the government should adopt or adapt to shape those laws.

Officials and some MPs argue that existing regulatory tools are not designed for the unique properties of large machine learning models, including their rapid release cycles, ability to autonomously probe remote systems, and dependence on concentrated compute resources. By insisting on independent evaluations and local compute allocations, the committee aims to create enforceable levers that can be applied when foreign-owned frontier labs want to train or run large models in Australia.

Procedural dispute over draft edits

The committee’s letter has not been without controversy. Independent senator David Pocock raised concerns that a senior adviser in the prime minister’s office reviewed and suggested edits to the bipartisan letter before it was sent. Documents seen by reporters show tracked changes and signoffs from a ministerial adviser on parts of the text, prompting objections from crossbench members who said such vetting undermines independent committee scrutiny.

Supporters responding to the revelation argue that executive office input on technical drafts is a normal part of parliamentary engagement with government, and that the substance of the recommendations remains the product of elected committee members. Critics say the appearance of preclearance matters because the letter frames early public policy positions that will influence draft legislation and industry negotiations.

Industry and civil society reactions

Early responses split along predictable lines. Security specialists and some national security voices welcomed provisions for early model access and independent testing, saying they would help surface risks to critical infrastructure and national data. Public interest groups and unions welcomed stronger reporting duties, but warned that procedural shortcuts risk undermining public trust.

Creators and parts of the media sector expressed alarm that committee language could presage changes to copyright or data access arrangements, which several stakeholders have argued must be handled with care. The committee’s letter did not spell out specific intellectual property reforms, but it stressed the need to consider licensing and remuneration where domestic content is used to train models.

Why the measures matter for Australia

The committee’s recommendations aim to reconcile two competing national objectives. On one side, Australia wants to attract investment and the capability benefits that come with large-scale AI development. On the other side, policymakers are concerned about the concentration of model training and inference offshore, the difficulty of enforcing safety practices across borders, and the risk that Australian data will be exposed or misused without timely notice.

Requiring independent safety evaluators and reserving compute capacity are practical policies that seek to make those objectives compatible. If implemented, they would give regulators a clearer sightline into how models are built and tested, and preserve a measure of domestic capacity to investigate incidents and to support public sector uses that may be overshadowed by commercial priorities.

What happens next

The committee intends to produce a fuller report later this year, with a formal reporting date set for 30 November. The government has said it aims to publish draft AI standards and related legislation by the end of 2026. The procedural dispute over edits is likely to feature in the committee record, and it may influence crossbench willingness to support recommended powers in draft law.

For businesses and research institutions, the immediate implication is greater policy uncertainty coupled with a clearer signal that Canberra is prepared to attach operational conditions to access and use of advanced AI resources. That combination could change negotiation dynamics when foreign labs seek to host training runs or data on Australian soil, and it is likely to shape commercial arrangements for compute and licensing in the months ahead.

As Australia moves from consultation to lawmaking, the balance policymakers strike between attracting investment and enforcing safety will determine whether the country becomes a host for frontier AI activity, or primarily a consumer of offshore capabilities subject to tighter domestic guardrails.