AI is shrinking the cyber defence window

On 22 June, the leaders of the Five Eyes cyber security agencies released an advisory regarding the “AI shift in cyber risk: why leaders must act now”.  

This is highly unusual. Its urgency, its call to action and its sign-off by the leaders of these organisations are rare.

The central premise is this: “Frontier AI models are anticipated to exceed current industry expectations, fundamentally transforming both offensive and defensive cyber capabilities. The timeline is not years, it is months.”

Why is this occurring and what can we practically do about it?

Hype or reality

Much has been made of the release, or lack thereof, of the Anthropic Mythos 5 and Fable 5 models, which are supposed to surpass all past models at finding vulnerabilities in software, such that all our businesses will be hacked in the near future.  

Anthropic, which is looking at an IPO soon, has made so much noise about this that the US government has banned these models from being seen by anyone who is not a US citizen, even if they work for Anthropic.

The Five Eyes advisory has been challenged by some; one futurist suggested it is a “triumph of pre-IPO marketing strategy”.  

Kordia’s specialist consultancy, Aura, and its team of ethical hackers and penetration testers have spent a lot of time researching the effectiveness of earlier Anthropic models (Sonnet and Opus) for finding and exploiting vulnerabilities in software.

In expert hands, these tools are very impressive – sometimes doubling the number of impactful vulnerabilities we would normally find within a timeboxed testing window.

In a penetration test, we find this kind of vulnerability through the knowledge and intuition of a skilled consultant directing the high-volume search and pattern matching capability of a large language model.  

The same capability also collapses the time available for an organisation to patch its software.

In the real world, a penetration tester or security researcher finds a vulnerability in a common software application or library. The vendor of that software creates a patch and makes it available to download. In the past, defenders would patch their software about once a month. In the meantime, attackers would reverse-engineer that patch to find out what vulnerability it was fixing, then attack that vulnerability. Hopefully, by the time they were able to attack the vulnerability, defenders had patched their software and it was no longer vulnerable.

With AI, the time to reverse-engineer patches has collapsed to an average of five days, and it is expected to get faster. Patching monthly is no longer viable.

Five Eyes advice

The Five Eyes advisory sets out five practical actions. They are genuinely practical, but they will also take time and may carry a cost, so they need timely executive attention.

Four are operational information technology and security activities your team can own: reduce your attack surface by cutting unnecessary network exposure, address legacy systems, strengthen identity and access controls, and accelerate your patching.

The fifth is different: prepare for incidents before they happen. This belongs squarely on the executive agenda because it assumes a breach will occur and tests whether your organisation can recover from it.

This last action is where the gap shows. The advisory recommends testing response plans regularly. Yet, according to Kordia’s 2025 Cyber Security survey of 250 New Zealand businesses with more than 50 seats, only half have done so in the past year. That means half of those New Zealand businesses are not prepared for this new paradigm.

The Five Eyes agencies also note that integrating AI tools into security operations can enable more effective detection of vulnerabilities and anomalous behaviour, as well as faster response to attacks.  

In practice, this means software such as CrowdStrike, which updates multiple times a day, can automatically isolate a machine when it detects a potential attack and give ethical hackers (and cyber criminals) untold grief when trying to hack into a network.

The changing paradigm

I believe we are sailing into a perfect storm in cybersecurity and that we need to do more than this.  

Coupled with the velocity of vulnerability exploits, companies are deploying AI and AI agents, creating a new attack surface while adversaries are becoming more audacious. The New Zealand National Cyber Security Centre (NCSC) reported three “highly significant incidents” in the first quarter of this year – the last one was five years ago.  

The New Zealand government is proposing to bring our critical infrastructure and privacy legislation into line with the rest of the world, which means companies and directors could face penalties if an incident occurs and their security measures are found to be inadequate.

For organisations coming to grips with the ever-increasing rate of change, the first place to start is a security assessment: how exposed are your critical assets and processes, and what would reduce that risk?  

If your organisation already has a vulnerability management programme, it is important to minimise your attack surface and maintain a patching cadence that reflects the evolving environment. 
The board should have sufficient visibility of this exposure, so it knows whether the  organisation’s critical assets are protected.  

I close by quoting the Five Eyes advisory one more time:

“Success will not come from having the most tools. It will come from getting the basics right, acting quickly, and integrating cybersecurity into core business strategy.”