Mythos travels to Tokyo: Japanese banks to acquire Anthropic’s AI for identifying vulnerabilities.

      MUFG, Mizuho, and SMFG are set to be the first Japanese banks included in Anthropic’s limited Project Glasswing rollout, a source familiar with the situation revealed to Reuters. According to the source, Japan's three major megabanks will gain access to Claude Mythos, Anthropic’s AI model for identifying vulnerabilities, in approximately two weeks. This marks the first occasion that a Japanese entity has been allowed into the restricted preview, which has previously been limited to Anthropic’s partners in the U.S. and a select few in Europe.

      During meetings in Tokyo this week with U.S. Treasury Secretary Scott Bessent, Mitsubishi UFJ Financial Group, Mizuho Financial Group, and Sumitomo Mitsui Financial Group were informed of this development. The three banks are anticipated to be onboarded by the end of May.

      Since Anthropic announced the model earlier this month, regulators and executives have viewed Mythos as a significant development. The AI model has identified thousands of previously unknown zero-day vulnerabilities across all major operating systems and web browsers, and internal tests revealed that it was capable of writing working exploits, including those that can bypass both renderer and operating-system sandboxes in a browser. Last week, Mozilla released Firefox 150, fixing 271 vulnerabilities detected by Mythos in a single evaluation session.

      Anthropic has not made the model publicly available but is conducting a controlled rollout under Project Glasswing, which includes 12 designated launch partners, such as AWS, Apple, Cisco, Google, JPMorgan Chase, Microsoft, Nvidia, and Palo Alto Networks, with another 40 institutions granted access on a selective basis. Japan's entrance comes shortly after the Federal Reserve and U.S. Treasury convened discussions with U.S. bank CEOs regarding cyber risks, and following commitments from UK regulators to brief major British banks promptly.

      Meanwhile, Tokyo is advancing in parallel. Finance Minister Satsuki Katayama announced the creation of a public-private working group consisting of 36 entities focused on Mythos-class risks, featuring the nation’s leading banks, the Bank of Japan, and the Japanese branches of Anthropic and OpenAI. The group, led by Mizuho’s Chief Information Security Officer, is tasked with identifying vulnerabilities, implementing defensive strategies, and formulating contingency plans for a coordinated patching initiative across Japan’s financial sector.

      For the three banks, the immediate concern is operational. Access to Mythos under the terms of Project Glasswing comes with restrictions on output disclosure, meaning the model is used solely to identify vulnerabilities within a partner’s systems and to develop remediation strategies, rather than to publish exploits. The Mozilla example provides a reference: 271 vulnerabilities were fixed in a single Firefox release following a Mythos review, with the model's findings returned to Mozilla engineers under a non-disclosure agreement rather than being made public.

      The geopolitical implications are clearly evident. Bessent’s role in delivering the access decision in Tokyo aligns the Mythos rollout with U.S. Treasury diplomacy rather than Anthropic’s commercial strategy, a situation that has elicited complaints from European capitals. Eurozone finance ministers raised concerns about this at an Ecofin meeting last week, noting that no EU government had access to the model while reports indicated that the White House was blocking the expansion of the partner list.

      Opinions within the industry regarding Mythos are divided. Some cybersecurity researchers contend that the vulnerabilities uncovered by Mythos can be reached through clever manipulation of public models, suggesting that the key issue is the rapid advancement of frontier AI in offensive cybersecurity rather than Mythos itself. Others, including Anthropic CEO Dario Amodei, have labeled this period a "cyber moment of danger," justifying the access controls set in place. Anthropic and the three Japanese banks did not immediately respond to requests for comments, according to the source cited by Reuters.