4. Artificial Intelligence

Anthropic’s new AI model crosses the Rubicon, Raises Questions of Business Survival.

Gallery Image — Anthropic’s chief executive Dario Amodei was in Australia last week meeting government and business leaders.

Anthropic’s chief executive Dario Amodei was in Australia last week meeting government and business leaders.

Anthropic, Creator of Claude AI (one of the better writing tools currently on offer) has acknowledged it is testing a new AI model that represents a ‘step change’ in capabilities. It had no choice really; an accidental data leak brought Mythos to the world’s attention.

Alastair MacGibbon - the founding head of the Australian Cyber Security Centre and the Special Adviser to the Prime Minister on cybersecurity – last week in the Australian Financial Review called upon the Cybersecurity Minister to convene critical infrastructure operators, private sector defenders and Artificial Intelligence developers - within a week!
He had met with Anthropic’s US leadership and learned how alarmed they are by what they have built.
The company’s unreleased Claude MythosPreview AI model can chain together security vulnerability flaws autonomously.
Claude Mythos has identified thousands of ‘Zero Day Vulnerabilities’ (look it up) across major operating systems and browsers.
What previously required rare human talent and expensive hacker tradecraft has been commoditised overnight.
This doesn’t just raise the threat level; it breaks the risk management model organisations have relied on for decades.
Executing major attacks at scale previously required nation-state resources or rare human hacker tradecraft, but Mythos has destroyed that assumption.

The vulnerabilities Mythos is finding exist because decades of poorly written code sit inside systems never designed for this threat. The answer is not just better patching, it is better code and new ways of architecting technology.

Alastair MacGibbon, Special Adviser to the Prime Minister on cybersecurity has called for an urgent response

'We have crossed the Rubicon. Development cycles are now faster than institutional response. We will not be overwhelmed gradually. We will be hit everywhere, across everything, all at once.' Alastair MacGibbon

Enter ‘Project Glasswing’

Anthropic has named eleven USA company partners, plus another 40 whose identities have not been publicly disclosed, who have been given exclusive access to the Claude Mythos Preview. The aim? To provide leading technology organisations with a chance to address the worst of their vulnerabilities before it’s too late.

Although access to Claude Mythos is currently restricted, it will not stay that way. Even if Mythos stayed contained, Open AI’s models are already strong at vulnerability research and China’s frontier labs are not far behind.

In one case, Claude Mythos identified a 27 year old vulnerability in a system that underpins the security of millions of organisations.

Remember how a faulty CrowdStrike update in 2024 grounded airlines, paralysed hospitals and severely disrupted financial markets, including at organisations that didn’t run CrowdStrike? This was because their suppliers (or their suppliers’ suppliers did), and nobody had mapped the chain.

That was an accident but imagine Mythos-class capability applied to the same interconnection by a criminal group or rogue state. For critical infrastructure owners and operators, Alastair MacGibbon describes the situation as a question of whether our society continues to function.

In his view ‘we have crossed the Rubicon’. ‘We will not be overwhelmed gradually. We will be hit everywhere, across everything, all at once’. Anthropic has called for regulation, but regulation always lags capability by several cycles.

What should all First National principals do?

Remind your staff regularly of cybersecurity risk and do everything possible to maintain high levels of awareness. We haven’t seen anything yet when it comes to the levels of scamming and hacking sophistication headed our way.

AI tells tenant she should ask for $40,000 - tribunal hands her $80

A New Zealand property manager has flagged a concerning trend: tenants using AI tools to prepare Tenancy Tribunal applications, often with inflated expectations and poor outcomes.

A New Zealand property manager described one case where a tenant, guided by AI, lodged a $40,000 claim over unsafe drinking water, alleged quiet enjoyment breaches and a broken dryer. The submission ran to 215 pages. The Tribunal awarded her $80.

He says AI-generated claims are arriving in increasing numbers, formatted identically and seeking tens of thousands of dollars. The problem is not that tenants are asserting their rights. It is that AI is providing unverified guidance to people who lack the legislative knowledge to question it, producing bloated, meritless claims that burden landlords, property managers and Tribunal adjudicators alike.

Tenancy Advisory director Sarina Gibbon offers a more balanced view, arguing that AI can genuinely broaden access to justice. The issue, she says, is that most users are still too unfamiliar with AI's limitations to use it well.

New Zealand's Insurance and Financial Services Ombudsman raised identical concerns last week, after AI tools told insurance complainants that 80 to 90 per cent of escalated cases succeed. They do not.

The lesson for Australian and New Zealand estate agents is straightforward. What plays out in New Zealand's tribunals today has a habit of appearing in ours shortly after. AI literacy, on all sides of a tenancy dispute, is no longer optional.