Japan’s megabanks to gain access to Anthropic’s vulnerability‑hunting AI Mythos

Japan’s three megabanks—Mitsubishi UFJ Financial Group (MUFG), Mizuho Financial Group and Sumitomo Mitsui Financial Group (SMFG)—are slated to receive Anthropic’s Claude Mythos AI model within roughly two weeks, a source familiar with the matter told Reuters on Tuesday. The banks will be onboarded by the end of May, marking the first time a Japanese institution joins Anthropic’s tightly controlled Project Glasswing preview.

Project Glasswing, Anthropic’s limited rollout, currently includes twelve named launch partners such as Amazon Web Services, Apple, Cisco, Google, JPMorgan Chase, Microsoft, Nvidia and Palo Alto Networks. Around forty additional organizations have been granted access on a case‑by‑case basis. Anthropic has kept Mythos out of the public domain, allowing partners to run the model under strict nondisclosure terms that limit output to internal vulnerability discovery and remediation.

Mythos has already demonstrated a capacity to uncover thousands of previously unknown zero‑day vulnerabilities across major operating systems and web browsers. In internal testing, the model generated working exploits, including chains that escape both renderer and operating‑system sandboxes in browsers. Mozilla’s recent release of Firefox 150 incorporated fixes for 271 vulnerabilities identified in a single Mythos evaluation pass, illustrating the model’s practical impact.

The decision to extend access to Japan emerged from meetings in Tokyo that included U.S. Treasury Secretary Scott Bessent. His involvement ties the rollout to U.S. statecraft rather than purely commercial channels, a point that has drawn criticism from some European capitals. Eurozone finance ministers raised the issue at an Ecofin meeting, noting that no EU government currently has access to the model while the White House reportedly blocks further expansion of the partner list.

Tokyo’s response runs in parallel with the rollout. Finance Minister Satsuki Katayama announced the creation of a 36‑entity public‑private working group to address Mythos‑class risks. The group brings together the three banks, the Bank of Japan and the Japanese units of Anthropic and OpenAI. Chaired by Mizuho’s chief information security officer, the consortium will identify exposures, implement defensive measures and draft contingency plans for a coordinated patching effort across the Japanese financial system.

For the banks, the immediate challenge is operational. Under Glasswing terms, Mythos can be used only to locate vulnerabilities in a partner’s own environment and to draft remediation steps; public disclosure of exploits is prohibited. The Mozilla example offers a template: findings are handed back to engineers under nondisclosure, enabling rapid patching without exposing details to potential adversaries.

Industry reaction remains split. Some cybersecurity researchers argue that the vulnerabilities uncovered by Mythos could be replicated through clever orchestration of publicly available models, suggesting that the broader story is the accelerating offensive capabilities of frontier AI. Anthropic’s chief executive Dario Amodei, however, describes the development as a “cyber moment of danger” that justifies the strict access controls.

Anthropic and the three Japanese banks did not immediately respond to requests for comment, according to the Reuters source. The rollout underscores a growing convergence of AI innovation, cyber risk and geopolitical considerations, as regulators and policymakers grapple with how to balance rapid technological progress against the need for robust defensive safeguards.