Earlier this year, the head of the Australian Cyber Security Centre warned that frontier AI is accelerating how quickly adversaries can find and exploit internet-facing systems. They’re right to sound the alarm: the Australian Signals Directorate (ASD) responded to more than 1,200 cyber security incidents in 2025, an 11 per cent increase on the previous year, and received an average of one cybercrime report every six minutes.
There’s no doubt cyber defenders are feeling the pressure, and for many Australian security leaders, “AI resilience” is the key priority. But “AI-resilient” runs the risk of becoming another corporate buzzword that looks impressive on paper while meaning very little in practice. What does it actually involve?
I anticipate the most resilient programs will be engineering-led and outcome-based; these programs focus on building scalable, repeatable solutions and flexibility to help the program adjust as AI technology advances. Here are three things every Australian organisation needs to get there:
Your team needs a culture of curiosity
Adopting a new technology (and doing it well) demands that everyone on your team gets curious; it’s a catch-all characteristic that cultivates the right actions. It encourages teams to find new ways to approach challenges, learn how things work, and produce outcomes outside their intended use case. Curiosity naturally grows an effective adversarial lens, something that hones how your teams identify AI risk.
This encourages teams to shift away from being reactive maintainers and advisers. They need to see themselves as system designers and builders of scalable solutions that automate outcomes for the business, informed by their understanding of what could go wrong.
To build this kind of culture, incentivise the right things. Make sure you’ve built structured learning and upskilling into how you support your team, encourage experimentation, and reward them for it. ASD Essentials already gives Australian organisations ready-made structure and materials to start learning.
Your tool stack needs flexibility
When you buy tools, look for solutions that ensure the choice of model and deployment stays with you, within your own guardrails. If you have strict data residency and classification requirements, you should also look for tools that give the option to run models in air-gapped or self-hosted environments.
Considering both frontier and open-weight models further adds to that flexibility, giving you more choice and helping you customise for your specific program and budget. While frontier models lead on raw capability, open-weight models give teams more control over how they run and often come at a much lower usage cost.
This flexibility guarantees that, as frontier models advance, you won’t be locked in when the best AI tools for cybersecurity today are no longer right in six months. So if you want to swap something out, your team isn’t forced to rebuild existing workflows.
Your agents need guardrails and governance
Trust in an agent’s output starts before the agent runs. It’s worth defining what the agent may touch and what a successful result looks like in measurable terms. Record tool calls and approvals, and tie each event to the final outcome delivered by each agent.
Traceability and permissions are very important; your team needs to see what each agent touched and which approvals it received as the work happens, rather than reconstructing events after an incident. Each agent should hold its own identity and bind every action to the human who set it in motion.
The reality is, policy has to live at the point where the agent acts. A rule that sits in a document, or a review that happens after a change ships, finds problems once the damage is done. Give agents a narrow scope so your team can better control an agent’s output and limit the damage if an agent goes off course. Enforce permissions, approvals, and separation of duties inside your workflows, so agents hit a guardrail before they touch sensitive information or a production system.
Our Global DevSecOps Report found that Australian teams trust AI to handle only 39 per cent of their daily tasks without human review. Teams that can show what an agent did, under whose authority and with what access, can extend that trust deliberately so agents can actually act autonomously, while human attention focuses on the decisions that need it.
Where enterprises are falling short
The biggest risk is analysis paralysis. Too many organisations are waiting for the perfect model, the perfect governance framework or an entirely new stack before they start.
It’s time to assess each area of focus, define the deterministic outcome you want to achieve, and work backwards using the technology already available. Then, iterate.
Nearly all security tools started out as a way to test and exploit environments, and security teams later repurposed them for defence. This logic applies here too; when teams that experiment gain an understanding of AI at a system-level, they’re much faster at finding new ways to build AI-powered defences.
An AI-resilient security function builds, tests, and adapts as fast as the technology does. Our adversaries are using the technology; it’s time for us to learn it too.
Want to see more stories from trusted news sources?Make Cyber Daily a preferred news source on Google.