AI That Cracks Bank Networks?

Anthropic's Claude Mythos has become a target of caution in the finance and security industries.

The model was assessed as having strong potential to find and exploit vulnerabilities in some operating systems and security environments, and in response, finance ministers, central banks, and major banks have discussed countermeasures.

Key points:

• Canadian Finance Minister François-Philippe Champagne said Mythos was discussed in depth at an IMF meeting.
• Barclays CEO CS Venkatakrishnan said vulnerabilities in systems need to be understood and fixed faster.
• Bank of England Governor Andrew Bailey also said it must be viewed very seriously from the standpoint of cybercrime risk.
• The U.S. Treasury Department recommended that major banks inspect their systems before public release.

Anthropic has not released Mythos to the general public; instead, it has made it available on a limited basis to AWS, CrowdStrike, Microsoft, Nvidia, and others, conducting security verification under Project Glasswing.

In independent testing by the UK's AI Security Institute, Mythos Preview was good at finding vulnerabilities in weakly defended environments, but the evaluation found it was not dramatically better than the previous Claude model (Opus 4).

The key point is not the performance itself, but the fact that a model at this level could expose weaknesses in banks, web browsers, and core IT systems on a large scale.