Japan's Government and Megabanks to Gain Access to Anthropic's Claude Mythos

Japan's Government and Megabanks to Gain Access to Anthropic's Claude Mythos

Japan has secured a seat at one of the most tightly guarded tables in artificial intelligence. On June 2, 2026, Finance Minister Satsuki Katayama confirmed that the Japanese government and a group of the country's financial institutions — widely reported to include the three megabanks MUFG, Mizuho, and Sumitomo Mitsui Banking Corporation — have been granted access to Claude Mythos, the frontier cybersecurity model from US developer Anthropic that has unnerved security agencies for its ability to surface software vulnerabilities no one else can see.

The announcement, timed in coordination with Anthropic and the US Treasury, makes Japan one of the first non-US governments confirmed to obtain access to a model the company had previously restricted to roughly 50 hand-picked partners. It also delivers, almost to the day, on a commitment US Treasury Secretary Scott Bessent made during the Japan-US Finance Ministerial Meeting three weeks earlier.

"As had been agreed for some time between us and the U.S. Department of the Treasury, the participants include the Government of Japan as well as some Japanese financial institutions," Katayama, who also serves as minister of state for financial services, told reporters in Tokyo. "Secretary Bessent said that this would be done in about two weeks. We are now in the third week, and I believe that his commitment has been largely fulfilled."

A frontier model, deliberately gated

Claude Mythos, unveiled by Anthropic in April 2026, is the company's most powerful model to date for security work, described as capable of identifying thousands of zero-day vulnerabilities — flaws unknown even to the software's own makers — across a codebase over a period of weeks. That capability is precisely why the company kept it on a short leash, initially limiting Mythos Preview to about 50 organizations, including the US government and large American tech firms, to scan their own code before adversaries could weaponize the same techniques.

The June 2 expansion, part of Anthropic's "Project Glasswing" initiative, widens that circle to roughly 150 new organizations across more than 15 countries, bringing the total to around 200. Per Financial Times reporting cited by TechCrunch, the cohort spans US-aligned nations including Canada, France, Germany, India, South Korea, and Japan, and reaches beyond finance into power, water, healthcare, and communications. Named participants reportedly include Okta, Samsung, SK Hynix, NATO, and the EU's cybersecurity agency ENISA.

Anthropic framed the stakes bluntly. "What each partner has in common is that a successful attack on their codebase could be catastrophic," the company said in a blog post. "For most partners, we estimate that a major attack could affect more than 100 million people."

Why finance went first

Katayama was careful not to name individual banks, citing the private-sector nature of the contracts. Pressed repeatedly on whether the three megabanks were included, she demurred — "I will refrain from listing them" — but pointed reporters toward the obvious inference, noting that when she first discussed the matter publicly the relevant Glasswing participants "were mainly the three megabanks," institutions that had themselves described the threat as imminent.

The minister's reasoning for prioritizing finance was structural. Banking, she argued, is inherently cross-border and depends on enormous volumes of near-instantaneous Japan-US transactions, placing it among the first sectors exposed to AI-enabled attack. The dominant risk, in her telling, is economic crime rather than espionage. "If this kind of capability were to be used for some malicious purpose, the largest proportion would likely involve money-related issues, such as proceeds of crime, money laundering, or fraud," Katayama said, adding that maintaining cybersecurity "at least on par with those of leading financial institutions in the United States" is now essential to the credibility of Japan's banks.

Asked whether access had actually been turned on, she was unambiguous: "I understand that access has been granted, and I believe that is also consistent with the wording used by the U.S. Treasury." Actual testing, run under a Japanese version of the Glasswing program and involving the banks' IT vendors, would follow on a separate timeline. No government money changed hands — the deployment rests on commercial contracts between the banks and Anthropic.

Sovereign AI, security diplomacy, and Anthropic's Asia push

The deal sits at the intersection of three trends. The first is sovereign AI adoption, in which states treat frontier-model access as a strategic asset negotiated at the cabinet level. The second is the financial sector's pivot to AI-native cyber defense — Japan is running an OpenAI track in parallel, with Katayama confirming access to a comparably capable GPT-5.5. The third is Anthropic's own trajectory: the Mythos expansion landed a day after the company filed confidentially for an IPO, following a $65 billion raise at a valuation approaching $1 trillion. Extending its most sensitive model to allied governments positions Anthropic as an instrument of US technology diplomacy even as it courts public markets.

That alignment carries an obvious tension. A model designed to find catastrophic vulnerabilities is, by definition, dual-use. Katayama acknowledged the difficulty the US side faced in coordinating access given "classified information and defense-related matters" and "the earlier disclosure concerning access for the Five Eyes." Distributing such a tool to roughly 200 organizations widens the defensive perimeter but enlarges the surface across which the techniques could leak — a gamble Anthropic justifies by arguing rival models are coming regardless. OpenAI has already released a cybersecurity-focused GPT-5.5-Cyber to testing partners.

What to watch

The headline access is settled; execution is not. Watch for when the megabanks actually begin Mythos testing under Japan's Glasswing framework, and whether scope widens to the energy, water, and telecom operators Katayama said the government had flagged on economic-security grounds. Also worth tracking: whether other governments now demand the same access, and whether Anthropic's gated-distribution model holds as the partner list grows. For now, Japan's financial system has moved to the front of the line — whether that yields durable defense or simply a faster arms race is the next chapter.