The ripple effects of Claude Mythos Preview have now reached the financial sector. After Anthropic unveiled its new model earlier this week — and simultaneously declared it too powerful for public release — banks and regulators are responding on both sides of the Atlantic.
Canada Moves First
In Canada, the Canadian Financial Sector Resiliency Group (CFRG) — a body made up of bank representatives and regulatory authorities — met on Friday to discuss the risks posed by Claude Mythos. According to Anthropic, the model can find vulnerabilities in virtually every major operating system and web browser using relatively simple prompts. For banks that depend on exactly this infrastructure, that’s not an abstract threat.
The Fed and Treasury Step In
The United States is taking the issue seriously too. The Federal Reserve and the Treasury Department sat down for a meeting to discuss the heightened cyber risks from Mythos. The goal: making sure banks are aware of the potential dangers and adjusting their security measures accordingly.
Project Glasswing as a Countermeasure
Anthropic itself launched Project Glasswing, an initiative where Mythos Preview is being used specifically to find and patch critical software vulnerabilities — before attackers can exploit them. Access to the model is currently limited to selected tech companies including Amazon, Google, Apple, Microsoft, Nvidia, and Cisco.
What This Means
The situation is unusual: an AI company develops a model, declares it too dangerous for the public, and triggers regulatory activity in the process. That’s new — and it shows that the conversation about AI safety is no longer confined to research papers. It has arrived in the conference rooms of central banks.
Whether this leads to concrete regulation or stays at the warning stage remains to be seen. What’s clear is that Claude Mythos has sparked a debate that extends well beyond the tech industry.
Sources: BNN Bloomberg: Anthropic’s new AI model is too dangerous to release · CNBC: Anthropic Claude Mythos AI hackers cyberattacks