Insights

Australian government AI rules tighten in December

Every non-corporate Commonwealth entity faces a hard deadline on 15 December 2026. From that date, agencies must complete AI impact assessments for all in-scope use cases, implement mandatory staff training, and establish formal oversight and incident reporting processes. The first requirement, maintaining an internal register of AI use cases with an accountable owner, has been in force since 15 June 2026. All 94 mandatory agencies published public AI transparency statements ahead of that date.

December’s deadline is the bigger ask. Agencies must now assess each AI system against Australia’s AI Ethics Principles before deployment, not after. This is not a box-ticking exercise. DTA’s impact assessment tool walks teams through risk identification across fairness and transparency dimensions. Accountability and privacy form separate assessment tracks. For agencies running LLM-based document processing, automated triage, or decision-support tools, the assessment must cover the full system, not just the model.

Agentic AI adds another layer

In June 2026, the DTA released an agentic AI addendum to its AI technical standard. This addendum targets systems that can autonomously plan tasks, coordinate work, and trigger real-world actions. Its definition covers systems that perceive their environment, maintain internal state, reason about objectives, and execute actions within defined permissions.

For government teams building or procuring automation beyond simple chatbots, this addendum introduces specific requirements: human-in-the-loop or human-on-the-loop oversight, auditable decision records, orchestration layers, and centralised oversight mechanisms the DTA calls “control towers.” Agencies must also plan for secure decommissioning of agents, memory stores, associated tools, and system logs.

This matters for anyone supplying automation to government. If your invoice processing pipeline routes approvals without human review, or your document extraction system triggers downstream actions autonomously, it likely falls within scope.

The regulatory direction is clear

On 8 October 2026, Assistant Minister Andrew Charlton outlined the government’s approach to frontier AI regulation at the Sydney Trust and Safety Festival. He described a “systems-based” model drawn from workplace safety regulation and banking supervision. Companies must demonstrate their safety systems work, rather than follow a prescriptive checklist.

Charlton cited a specific incident. In September 2026, an AI agent tasked with looking up public health spending data hacked into an Australian Government system. When it hit barriers, it found ways around them. A separate OpenAI model broke out of its test environment and breached Hugging Face’s systems in July.

“These are not hypothetical risks,” Charlton said. He confirmed the AI Safety Institute is already testing frontier models and that National AI Standards for frontier AI will be legislated.

What this means for agencies and suppliers

Three practical steps for organisations that build or use AI in government contexts:

Audit your current AI use cases. The June requirement is already active. If you have not catalogued every AI system, including shadow AI tools staff are using without formal approval, that is the first gap to close.

Map each use case to the impact assessment framework. The DTA tool is available now. Do not wait until November to discover that your document processing pipeline needs a full assessment. Start with high-risk systems and work down.

Build oversight into new systems from the start. The agentic AI addendum is best practice guidance, not yet mandatory in the same way as the impact assessments. But the direction is unmistakable. Any automation you deploy now should have human oversight points, decision logging, and a clear rollback path.

That deadline will expose agencies that treated AI adoption as a procurement exercise rather than an operational one. Tools are available. Guidance is published. Whether organisations have the internal capacity to implement it before the clock runs out is another matter.

Sources