Why these AI tools are deemed too dangerous to release

Saturday, 11 July 2026

Anthropic's powerful new AI model, Mythos, is once again accessible after being ordered offline by the White House to implement safeguards.

During testing, Mythos found thousands of high-severity vulnerabilities in core systems, including some in every major operating system and web browser.

The NCSC responded to 1,164 incident reports in Q1 2026, including three critical 'C2' incidents affecting sensitive data or essential services

While hard numbers are understandably hard to come by, estimates for the global cost of cybercrime are as high as $17 trillion NZD annually.

If it were a country, that means it would have the 3rd highest GDP worldwide, behind only China and the US.

That’s one reason governments worldwide are ringing alarm bells over how frontier AI tools, like those made by OpenAI and Anthropic, can supercharge cybercrime.

A toned down version of the latest release from Anthropic, Mythos, is now once again accessible to Kiwis after the White House ordered it taken down worldwide to develop additional safeguards.

Even the most secure IT infrastructure on Earth proved vulnerable to a full powered Mythos.

The Pentagon’s Head of Cyber Command reportedly said when given access to a test version, the model “broke into almost all of our classified systems, not in weeks, but in hours”.

The head of New Zealand’s Cyber Security Centre (NCSC) recently took the unusual step of joining our Five Eyes Allies in a public statement calling for greater urgency in responding to AI which is transforming cyber security in a matter of ‘months not years’.

“They're much better at reasoning and decision making than the previous generations of AI. This is now really, really competent with extremely technical data,” Deputy Director-General of Cyber Security for New Zealand, Catriona Robinson, told Stuff.

But what are these new AIs? How powerful are they? And what can we do in response?

Why the new models are raising alarm

Fundamentally, the same thing that makes an AI system useful is the same thing which makes it dangerous.

The better it is at studying viruses in a research setting, the better it is at potentially making a new bio weapon.

And the better it is at writing and understanding code, the better it is at finding vulnerabilities with that code.

Most of the AI tools we use, like the chatbots, are in some ways a side hustle for Silicon Valley.

The primary focus is making them better at coding so that AIs can increasingly build themselves. Software developers are expensive and work slowly compared to a bot.

This is sometimes called ‘recursive self improvement’, with Anthropic estimating its newest model wrote 80% of its own code. Whoever builds an improved model fastest pulls ahead in the AI race.

There’s also a chance this kind of self improvement kills everyone on Earth but that's a separate story.

The danger from a cybersecurity perspective is that much of the infrastructure our digital economy depends on is built using a relatively small number of systems.

During testing, Mythos found thousands of high-severity vulnerabilities in those systems, including some in every major operating system and web browser.

For example the Linux Kernel, a core program managing computer hardware, is used by 5.5 billion devices globally.

Linux underpins approximately 96% of the top million web servers and about 80% of all cloud computing infrastructure.

Some of the vulnerabilities Mythos discovered were decades old, with the oldest being a 27-year-old bug in OpenBSD—an operating system known primarily for its security.

Big Tech Response

The obvious danger of this led Anthropic to form Project Glasswing, an emergency coalition where it shared Mythos with multiple tech giants and governments, including New Zealand, so experts could locate and patch vulnerabilities before malicious actors can abuse them.

The US Government also stepped in with OpenAI, requesting it release its newest model first in preview form to select partners in a similar programme.

The primary risk comes from so-called ‘Zero Day’ exploits, where a hacker discovers a previously unknown vulnerability in the code of a website.

The owners of that code have no warning of a vulnerability, hence having ‘zero’ days to prepare.

An entire online economy exists to buy and sell zero day exploits - both legitimately and criminally. Governments and businesses can offer ‘bug bounty’ programmes for coders who reveal a flaw in their security systems.

But of course in the internet’s darker corners zero day exploits are bought and sold for millions.

The State of Cybercrime in NZ

Kiwi dollars are propping up this shadow economy.

Cyber security incidents surged in Q1 2026, costing Kiwis $5.6m in direct financial loss - a 76% jump from the previous quarter, according to data from NCSC.

Between January and March 2026, the NCSC responded to 1,164 incident reports, only slightly above the 1,131 reported to round out 2025.

While the overall number of incidents was similar, the impact was much greater.

The vast majority of loss was concentrated in just 42 incidents - making an average loss of over $100,000 in each case. Three incidents were categorised as C2 or ‘highly significant’. These are the first C2 incidents recorded by NCSC since 2021.

C2 incidents are defined as impacting key sensitive data or causing disruption to essential services in organisations of national significance.

Of 77 incidents requiring specialist support by the NCSC almost a fifth (17%) were assessed as likely linked to state-sponsored actors. Around half (52%) were linked to cybercrime actors, and 31% did not have enough evidence to link the activity.

What Kiwis should do

If you’re reading this and you’re worried about how to protect yourself, luckily, just because the tools are more sophisticated, the defences against them are often old school.

Tools like two factor authentication are still effective in the majority of cases as well as basic cyber hygiene like ensuring all of your systems are patched up to date.

For businesses, AI can also be a powerful tool for defence by scanning for issues to be patched before a bad actor can find them first.

And while the rest of us grapple with what AI could mean for the future of our jobs, even Robsinson is in the same boat, having assumed her role in only March this year.

“We might need to readjust how we think about cyber defence and the role of the NCSC in this new world. We're at the beginning of thinking about that and while I don't have a good answer yet, you'll understand it's top of mind for me,” she told Stuff.