Your AI tool is sending client data offshore. And you are legally responsible for it.

The CLOUD Act is not optional

The US CLOUD Act (2018) lets US federal authorities compel American technology companies to produce data they hold, wherever in the world it sits. It applies to the provider, not to the data's location.

This creates a situation most Australian organisations have not modelled: you can be fully compliant with the Australian Privacy Act and still have exposed client data to a foreign jurisdiction. Both obligations apply simultaneously, and satisfying one says nothing about the other.

The 10,000 model reality

The counter-argument used to be that you had to accept this trade to get frontier capability. That has stopped being true for most enterprise work.

There are now more than 10,000 open-source models available to download. Llama 3, Mistral, Falcon and Qwen are production-ready and commercially deployable, and over roughly eighteen months the performance gap has closed for the task types regulated industries actually run, summarisation, extraction, classification, drafting and retrieval over your own documents.

Running one of those inside your own environment means no external API call, no data leaving, and no US jurisdiction attaching to any of it.

The frameworks that already apply

• Privacy Act, APP 8, cross-border disclosure of personal information
• APRA CPS 230 (from July 2025), governance of material service providers
• APRA CPS 234, information security, including third-party API risk
• My Health Records Act, restrictions on offshore processing
• NDIS framework, participant data protection
• ISM PROTECTED, classified data handling requirements

CPS 230 deserves particular attention

Banks, insurers and superannuation funds must formally govern their material service providers. In practice, most organisations using frontier AI tools have neither registered them as service providers nor run a risk assessment on them.

That is not a policy gap to tidy up later. It is a prudential compliance failure, and it is enforceable.

What sovereign AI architecture actually requires

1. The model runs on infrastructure located in Australia
2. No data leaves the environment during inference
3. Full visibility and control over model behaviour
4. No connection to an external training pipeline
5. Demonstrable alignment with the frameworks that bind you

All five, or it is not sovereign. Four out of five is a marketing claim.

How BlackVault™ delivers it

• Model selection matched to your specific use cases
• Sovereign deployment on your infrastructure
• Domain fine-tuning on your organisation's data
• Compliance architecture with audit logging built in
• Ongoing operations management, permanently