Finance ministers from 14 countries held crisis talks at this week International Monetary Fund meeting over one artificial intelligence model they fear could crash the global financial system.

Anthropic Claude Mythos — the first AI model ever restricted from public release due to its destructive cybersecurity capabilities — was "discussed extensively" during the IMF meetings in Washington DC this week, according to Canadian Finance Minister François-Philippe Champagne.

"Certainly it is serious enough to warrant the attention of all the finance ministers," Champagne told the BBC. "The difference is that the Strait of Hormuz — we know where it is and we know how large it is... the issue that we are facing with Anthropic is that it is the unknown, unknown."

The model has already exposed "multiple security vulnerabilities in some critical operating systems, financial systems and web browsers," according to Anthropic. But its real power lies in what security experts call "exploit chains" — the ability to link vulnerabilities together and execute full system takeovers.

The Council on Foreign Relations warns that Mythos represents a fundamental shift in cybersecurity: "The ability to identify zero-day software vulnerabilities previously resided only with highly specialized and knowledgeable cybersecurity experts and hackers. No more."

Anthropic engineers "with no formal security training" could ask Mythos "to find remote code execution vulnerabilities overnight" and wake up "the following morning to a complete, working exploit," according to the company report.

Barclays CEO CS Venkatakrishnan told the BBC the model creates vulnerabilities "for security for the entire banking system." Bank of England Governor Andrew Bailey warned that "cyber criminals — the bad actors — could seek to exploit" the model capabilities.

The US Treasury has already briefed major banks, encouraging them to test their systems before any public Mythos release. Financial industry sources indicated that another prominent US AI company (likely OpenAI) could soon release a similarly powerful model "without the same safeguards."

Rather than release Mythos publicly, Anthropic created an elite consortium called Project Glasswing, giving early access to Amazon, Apple, Google, Cisco, JPMorgan Chase, Microsoft, and Nvidia. OpenAI is conspicuously excluded from the consortium.

"Every leap in AI capability doesn expand the attack surface itself, chaining flaws in ways we have not imagined," Lucas Nelson of Lytical Ventures told CFR. "Discovery is accelerating exponentially. Remediation still moves at human speed. That asymmetry is the defining cybersecurity challenge of the next decade."

The model has already found vulnerabilities in systems "10 or 20 years old, with the oldest being a now-patched 27-year-old operating system." In one case, Mythos found a flaw in code that had been tested five million times without detection. Anthropic said 99% of the thousands of zero-days it discovered remain unpatched.