While AI continues to make the news globally, including the reported incident in which an OpenAI agent reached an Australian Medicare statistics portal, the Australian Public Service (APS) is looking at how to implement and scale AI within its operations to improve productivity.
The lesson from incidents like it is not that AI creates an entirely new category of risk, but that it can amplify familiar ones: misconfiguration, excessive access, and insufficient understanding of where sensitive data sits. Gaps that once took human time and effort to find can now be discovered far more quickly.
How can this be done safely to prevent risks and build trust?
According to a new whitepaper, Taking a Measured Approach to AI in the APS, building resilient infrastructure is critical in combating potential risks to AI systems and requires all agencies to ask three key questions to assess their capability:
1. Can the agency recover the critical services and data AI depends on cleanly and in the right sequence?
2. Can the agency demonstrate recovery will work before an incident occurs?
Essential reading for Australia’s public service.
Stay ahead of policy shifts, leadership moves, and the big ideas shaping the public service. Sign up to The Mandarin’s free newsletters.
By continuing, you agree to our Terms & Conditions and Privacy Policy.
3. Does the agency retain sufficient control over data, identity, and recovery as their technology changes?
The third question is the one that has changed most. The whitepaper points to recent restrictions on international access to advanced AI models, a reminder that control over a system an agency depends on may sit offshore, and can be withdrawn on a political decision rather than a technical one.
The challenge is that the capability to deliver this resilient infrastructure may not currently exist in the APS.
Capability where it matters is low
According to the whitepaper, Essential Eight maturity is critically low within the APS. The Essential Eight is designed to help organisations protect their internet-connected networks and mitigate the risks of cyberattacks. This will be updated to “Essentials for Enterprise IT” over the next 12-to-24 months to deal with cyber security in the modern world. The refreshed framework prioritises stronger operational resilience, data protection, and recovery readiness.
Almost all strategies and controls in the “Essentials for Enterprise IT” are areas APS agencies are recommended to improve. Rob Johnston, Senior Systems Engineer at Veeam, advises agencies to question how good their practices are and to re-think strategies to support the current and new initiatives.
“Are backups immutable? Are they isolated/air-gapped? Are restores tested and verified?” he asked. Johnston says these questions will determine how resilient an agency will be in the event of an attack and how exposed it may be as a result of implementing AI solutions. “There is a clear connection between AI enablement and operational resilience,” he said.
Data quality and robust data recovery processes are prerequisites for trustworthy AI outputs.
Partnerships will improve outcomes
Data resilience and data security companies such as Veeam help the APS deliver safer AI solutions by ensuring the data and services AI depends on remain recoverable, controlled, and resilient, building public trust in the process.
“We’re focused on protecting the data, but our expertise in recovery is essential to the APS and risks AI brings,” Johnston said. “Our data security practices wrap powerful data security frameworks around the data to provide protection as well as intelligence as to who has access including AI models and where data is going.”
In the event of a major outage, services of providers, such as Veeam, ensure data can be lifted and recovered elsewhere. This data assurance provides a safety net for APS agencies in their AI journey that may not exist in-house.
The results are strong and trustworthy systems
Doing AI right is important to build community trust. The OAIC’s Australian Community Attitudes to Privacy Survey, released in May 2026, reveals Australians are cautious about the use of AI – particularly where decisions impact them. 71% of respondents were “somewhat or very uncomfortable” with personal data provided for a particular service being used to train AI, and 75% did not approve the use of AI to make decisions such as loan or benefit approvals.
Lack of control of personal data and information was a huge concern for Australians who increasingly felt the ability to control their information was diminishing with advancements in technology.
But taking a measured approach to AI and ensuring strong and trustworthy systems are built may change the public’s perceptions.
To find out more about building resilient AI in the APS, download the whitepaper today.