Sprinkling AI pixie dust around the place doesn’t automatically provide cyber defence. In its Annual Cyber Threat Report 2025–26, the Australian Signal Directorate urges organisations to embrace AI for cyber defence, but asking how much AI your cyber team is using is asking the wrong question.
No AI product makes up for not knowing what software you run, which of it is out of date, or what data you have needlessly exposed to the internet.
The ASD responded to more than 1,500 incidents last year, up more than 20 percent on the previous year, as attackers used automation to exploit known weaknesses faster than poorly managed organisations can fix them.
Boards and senior executives should re-evaluate their approach to cyber defence – but they must judge cybersecurity investments by measurable risk reduction rather than by the amount of defensive AI that’s been employed. Specific tools need to be matched to specific weaknesses. It is that question that should start every security decision: what is our biggest risk, and what is the best way to reduce it? Sometimes the answer will involve AI. Often it won’t.
If you read the ASD’s 26 October report in detail, you’ll see nuanced guidance about what is right for organisations of different scale and levels of maturity. But the well-intentioned headlines about it run the risk of being misread as an instruction to ‘just use AI’, while encouraging vendors to shoehorn ‘AI-powered’ into their marketing, whether it is relevant or not.
‘AI cyber defence’ is not one thing; it’s a variety of potential approaches, and the right tool (if any) varies by organisation. AI that automatically checks a system’s security settings can be useful for organisations with poor IT configuration management. A software developer with a robust quality assurance process can get real value from AI that checks source code for vulnerabilities as part of its development pipeline; most other organisations have no use for it. Agentic AI that detects and responds to incidents autonomously is likely to provide a net benefit only for organisations with very mature processes, well-understood networks and strong governance of AI itself.
Understanding the context of AI tool use and how it is governed is critical because these tools fail in ways that are hard to spot. Large language models hallucinate, producing confident findings about vulnerabilities that don’t exist. Open-source software maintainers have complained publicly about floods of bogus AI-generated bug reports. An organisation without the expertise to check outputs will waste scarce effort chasing phantoms – or, worse, trust a tool that misses real problems. Further risks can come from feeding security data to an AI service without controls on where this data goes and who has access to it. And an agent given broad permissions to act on a network can make mistakes at machine speed.
As attackers make more use of AI, the first and most effective step organisations can take is to get the basics right: strong authentication, installing security updates quickly, disconnecting systems that don’t need to be on the internet and making sure default settings are secure.
Some of this can be automated, but automation long predates the current AI boom. Automated patch deployment, attack surface scanning and configuration baselines enforced by policy are established, predictable technologies. They directly address the weaknesses attackers are exploiting without necessarily needing AI. The ASD report cites an example of attackers successfully exploiting a vulnerability first identified in 2008. No AI would have been needed to fix that: regularly installing software updates would have been enough. Similarly, the decades-old Medicare statistics portal recently accessed by an OpenAI agent didn’t need smarter defences. It just needed to be taken off the internet.
For boards and senior executives, the practical test is simple. Before approving any AI security investment, they should ask what specific risk it reduces, and whether a cheaper, more predictable control would reduce it better. If they do decide to consider buying ‘AI-powered’ cyber security tools, they need to ask hard questions of the vendors. What exactly does the AI do, and what would the product do without it? How often are its findings wrong, and how is that measured? What data leaves our network, and where does it go? What permissions does it need, and what can it do without a human approving it? How are its models updated and tested, and for how long will the vendor support it? A vendor that can’t answer these questions clearly is selling a label, not a capability.
The ASD is right that AI is reshaping cyber security for attackers and defenders alike. But success is measured by how much risk is reduced, not by how much AI is being used. Boards that start with their biggest risks will find the places where AI genuinely helps. Boards that start by simply telling managers ‘we need AI’ will end up buying answers to questions they never asked – and will leave their real risks untouched.
Leave a comment