Warning: The ‘too dangerous’ AI knocking on New Zealand’s door
Sunday, 26 April 2026
Kevin Norquay is a senior writer for The Post and Sunday Star-Times. This is his weekly explainer.
EXPLAINER: Silicon Valley has a familiar marketing trick - build something extraordinary, then hint it might be too dangerous for the rest of us as a way to trumpet its power (and value).
This month, that label has been attached to Claude Mythos Preview, a cybersecurity-focused AI model from Anthropic. Why? And should New Zealand be concerned?
Mythos has quickly become the latest flashpoint in the global AI debate — drawing in tech companies, banks and regulators.
At its core, it’s a system designed to find weaknesses in digital infrastructure. In the right hands, that’s a powerful defensive tool. In the wrong hands, it’s a fast-track to sophisticated cyber attacks.
Read more:
Anthropic is an artificial intelligence company founded in 2021 by former researchers from OpenAI. It focuses on building safe, reliable AI systems - with an emphasis on making them more predictable and aligned with human intentions.
So when it says Mythos is dangerous, ears prick up. What sets Mythos apart is not the nature of the threat, but the speed and scale. It can identify vulnerabilities far faster than humans can fix them - a shift that could tilt the balance between digital attack and defence.
For banks, government agencies and anyone holding sensitive data, that imbalance matters.
Critics say tools like this “weaponise” AI reasoning, accelerating the offensive side of cyber conflict.
Others hear a different signal - that talk of something being “too dangerous to release” can also serve as marketing, building mystique while limiting access.
And access, in this case, appears tightly controlled. A small group of major players - including Apple, AWS and JPMorgan - reportedly have early exposure, raising questions about who gets to hold the advantage.
Regulators in the United States and United Kingdom are watching closely, citing national security concerns and potential risks to financial stability. So where does that leave New Zealand?
According to Albert Bifet, professor of AI and director of Te Ipu o te Mahara - the AI Institute at the University of Waikato, this is not a moment for alarm - but it is one for attention.
“In simple terms, this is not a completely new threat — it’s a stronger version of something that already exists,” he says.
“Cyber attacks are not new. What is new is that AI makes them faster, cheaper and easier. It can find weaknesses in software, generate attack code, and do this at a very large scale.”
That scaling effect is the real shift. It’s not just that the most capable actors become more powerful. It’s that many more people can operate at a higher level.
AI can:
* Scan systems for vulnerabilities at speed
* Generate malware or attack strategies
* Produce more convincing phishing scams
* Allow less-skilled actors to behave like experts
“The concern is not only very powerful actors,” Bifet says, “but many more actors gaining these capabilities.”
For New Zealand, the response is relatively clear - if not simple. As a highly digital economy, it is exposed. But it is not without options.
Bifet argues the country should invest in AI and cybersecurity skills, strengthen its digital infrastructure, work closely with trusted partners, and develop what he calls “sovereign AI” - the capability to build and understand its own systems rather than relying entirely on others.
“If we do not invest in AI, we risk falling behind economically and strategically,” he says.
AI, he suggests, is becoming a general-purpose technology - more like electricity or the internet than a niche tool. Its effects will be broad, and unavoidable.
He points to Quantum computing as a future risk. Quantum computing uses the rules of quantum mechanics to process information in a completely different way from normal computers.
It’s faster and more flexible. Quantum systems could eventually break current encryption methods, although that threat is still emerging. AI is not.
“AI is already here, and improving very quickly. Its impact is immediate,” Bifet says. “That’s why we need to act now, while also preparing for future risks like quantum.”
The real danger, he suggests, lies at the extremes. Ignore it, and vulnerability grows. Overreact, and innovation slows - along with independence.
“The right balance,” he says, “is to invest, build capability, and manage the risks carefully.”
UK-based Kiwi James Parr, founder of FDL.ai which builds AI for NASA and ESA (European Space Agency), is similarly cautious.
“Reasoning models like Claude collapse pretty rapidly when following long chains of thought,” he says.
“It’s like someone who’s really good at a pub quiz, but can’t run a business. (This is an effect called context compression and you should be aware of this the next time you use ChatGPT to plan a holiday.)”
However, it likely does scale the capacity and ingenuity of malevolent actors, Parr warns.
“So the threat of Mythos is real, and we should applaud Anthropic for calling it out - but it’s also intrinsic to the discussion we should be having about Frontier Models going forward.
“A danger we don’t talk about enough is how these tools are ‘hacking humanity’. Issues like context compression, model sycophancy, subject drift and automation bias mean we should be cautious, whatever the model.”
What do you think? Email sundayletters@stuff.co.nz. Please include your full name and address.