MUFG, Mizuho and SMFG would be the first Japanese institutions added to Anthropic’s restricted Project Glasswing launch, a source familiar with the matter told Reuters.
Japan’s three megabanks will gain access to Claude Mythos, Anthropic’s artificial intelligence model for searching for vulnerabilities, in about two weeks, a source familiar with the matter said. Reuters on Tuesday.
It would be the first time a Japanese company has been granted entry to the restricted preview, which until now has been limited to Anthropic’s American partners and a handful of European ones.
Mitsubishi UFJ Financial Group, Mizuho Financial Group and Sumitomo Mitsui Financial Group were informed of the move during meetings in Tokyo this week with US Treasury Secretary Scott Bessent. All three lenders are expected to come on board by the end of May.
The 💜 of EU technology
The latest rumors from the EU tech scene, a story from our wise founder Boris and some questionable AI art. It’s free, every week, in your inbox. Register now!
Mythos has been treated by regulators and CEOs as a category-changing event since Anthropic revealed its existence earlier this month.
The model has discovered thousands of previously unknown zero-day vulnerabilities on all major operating systems and web browsers, and in internal testing wrote exploits that work, including strings that escape both the renderer and the operating system in a browser.
Last week, Mozilla shipped Firefox 150 with fixes for 271 vulnerabilities found by Mythos in a single evaluation pass.
Anthropic has not made the model public. Instead, it has done a controlled implementation under what it calls Glass Wing Projectwith 12 launch partners named, including AWS, Apple, Cisco, Google, JPMorganChase, Microsoft, Nvidia and Palo Alto Networks, and around 40 other institutions granted access on a case-by-case basis.
Japan’s inclusion comes weeks after the U.S. Federal Reserve and Treasury summoned U.S. bank CEOs for the same briefing on cyber risks, and after UK regulators pledged to inform major British banks in a few days.
Tokyo advances in parallel. Finance Minister Satsuki Katayama announced the formation of a 36-entity public-private working group on Mythos-class risks, made up of the country’s major banks, the Bank of Japan and the Japanese units of Anthropic and OpenAI.
The group is chaired by Mizuho’s chief information security officer and is tasked with identifying exposures, implementing defensive measures and drafting contingency plans for what would amount to a coordinated patching effort across the Japanese financial system.
For the three banks involved, the immediate issue is operational. Mythos under Glasswing’s terms ships with restrictions on disclosing results, and the model is used to find vulnerabilities in a partner’s own systems and write fixes, not to publish exploits.
The Mozilla case offers a model: 271 vulnerabilities patched in a single version of Firefox after a Mythos sweep, with the model’s findings returned to Mozilla engineers under disclosure rather than published.
The geopolitical layer is unusually visible. Bessent’s role in relaying the access decision in Tokyo aligns the Mythos launch with the US Treasury’s statecraft rather than Anthropic’s commercial channel, an arrangement that has drawn complaints from European capitals.
Eurozone finance ministers raised the issue at an Ecofin meeting last week, where no EU government had access to the model while the White House was reported to be blocking further expansion of the list of partners.
Industry opinions on Mythos remain divided. Some cybersecurity researchers have argued that The vulnerabilities that Mythos revealed are achievable through clever orchestration of public models.and that the bigger story is the rate of improvement of frontier AI in the cyber offensive, not Mythos itself.
Others, including Anthropic CEO Dario Amodei, have described the moment as a “cyber moment of danger” that justifies access controls.
Anthropic and the three Japanese banks did not immediately respond to requests for comment, according to the Reuters source’s account.
