OpenAI: Press Conference By KATAYAMA Satsuki, Japan Minister Of Finance And Minister Of State For Financial Services

[Opening remarks:]

Minister)

I have just held a meeting with Mr. Kwon, CSO (Chief Strategy Officer) at OpenAI’s headquarters. Mr. Kwon explained today that OpenAI would, in coordination with the U.S. government, provide certain financial institutions in Japan with access to GPT-5.5-Cyber, OpenAI’s latest frontier AI model, as set out in a memorandum. Given the current understanding that frontier AI can pose a threat, I believe this is a significant opportunity, and a welcome development, from the perspective of strengthening cybersecurity at Japanese financial institutions. Following the Prime Minister’s instructions, I will work in close cooperation with Minister Matsumoto, Minister for Digital Transformation and Minister in charge of Cybersecurity, and in an integrated manner with the National Cybersecurity Office (NCO). We held discussions last week as well, and will continue to work in close coordination with other relevant organizations, using this development as an opportunity to enhance Japan’s overall capabilities and resilience in responding to frontier AI. That is all from me.

[Questions and answers:]

Q.

First, now that access to OpenAI’s model has been granted, how do you intend to disseminate the knowledge gained through this access to other financial institutions, electric power companies, medical institutions, and other infrastructure operators?
Second, as moves to grant access to frontier AI models continue, how do you view possible future access to Google’s latest AI models and other frontier AI models of this kind? Also, regarding Mythos, at last week’s post-Cabinet press conference, you stated that when Secretary Bessent was in Japan, he said that access would be granted to financial institutions within two weeks. Is it correct to understand that such access has in fact already been granted?

A.

Let me begin with the latter part of your question. With regard to Mythos, at the bilateral Japan-U.S. Finance Ministers’ Meeting, we were told that the Japanese government and trusted Japanese financial institutions would be granted access. However, when they say two weeks, it sometimes turns into four weeks, so in that sense, we are currently in contact through the Vice Minister of Finance for International Affairs and the Financial Services Agency, and I understand that matters are proceeding on that basis. That said, on the U.S. side as well, they are probably considering various matters, such as how to handle the executive order. In the case of Mythos, there is the question of what its relationship with the Department of War is. We asked OpenAI about this today as well, but it was not clear. We were told that, since OpenAI is not subject to constraints of that kind, this approach would be acceptable for a U.S.-developed frontier AI model, following coordination with the U.S. government. At this point, however, there may still be matters to coordinate or fine-tune with regard to Mythos. At that time, Secretary Bessent referred to three models: Mythos; GPT-5.5, the OpenAI model discussed today, although I understand that a GPT-6 model may be released at some point; and a new version of Gemini. No timing has been indicated for the new Gemini version, but it is apparently not too far off. He said that, as far as the U.S. government is currently aware, these three models are frontier AI. In other words, these models are considered to be so highly intelligent that ordinary human management and patching processes would not be able to keep pace. I believe that there is, in broad terms, an underlying approach of setting a threshold for trustworthiness and allowing access to entities that meet that threshold. We would like to work out the details through our daily contacts with the United States going forward.
As for disseminating the knowledge gained, as I discussed last week with Minister Matsumoto and Mr. Iida of the NCO, if this becomes a matter of system-level responses and this access can serve as one test case, vendors that work with major financial institutions will naturally also have relationships with other major institutions. Where there are common elements, those vendors will have relevant know-how. Therefore, assuming that such vendors become involved in this way, I believe we can gain knowledge about the level of preparedness that would be advisable. In that sense, such knowledge can certainly be disseminated appropriately. However, whether the same can be said of other sectors—where such private-sector relationships of trust do not currently exist—as compared with certain Japanese financial institutions where a certain degree of trust already exists between private-sector parties is not something we can really comment on, because the contracts themselves are between private-sector parties.

Q.

I have two questions. You explained this at the outset, but in today’s briefing by OpenAI’s CSO, did he say that access has been granted, or that it will be granted?

A.

My understanding is that access has been granted, and that work will now begin. OpenAI is constantly releasing new models and undertaking various initiatives, and it also has a Japan office. The financial institutions in question are among OpenAI’s customers.

Q.

Could you provide the specific names of those financial institutions?

A.

They are customers, and the matter is between private-sector parties, so I think it would be best to ask the relevant parties. That said, I believe they are major institutions of the kind you would imagine.

Q.

I have one more question. You mentioned GPT-5.5 this time, while there has also been ongoing discussion about access to Mythos. Could you explain how you view the fact that access to GPT-5.5 has been obtained? For example, is it as significant as obtaining access to Mythos?

A.

There is also talk that Mythos itself has already evolved further as it has been used. Versions will presumably continue to be updated day by day. Therefore, the fact that GPT-5.5, which the UK’s AISI has recognized as having exceeded a certain threshold, is now entering experimental testing using such access is meaningful. At the same time, Mythos presumably has some additional functions of its own, so if we separately obtain access to it at the next stage, that will also be meaningful in its own right. Rather than focusing too rigidly on where to draw the line, we have already convened a working group on AI-related threats, and most of the key stakeholders on the Japanese side are participating in it. In a sense, it is a Japanese version of Glasswing. If we feed that knowledge into the working group, I believe that, overall, our response capabilities will improve, regardless of which frontier AI model is involved.

Q.

Much attention has focused on Mythos. Is it correct to understand that this is an equally significant development?

A.

In any case, until now, we had not had access to frontier AI. In that sense, I believe it is meaningful.

Q.

Regarding your meeting earlier, it appeared to go beyond the scheduled time. Could you tell us, in your own words, which parts of the discussion were particularly lively in your talks with OpenAI, or which discussions or exchanges of views stood out to you?

A.

In my office, I have photographs from the Japan-U.S. summit meeting at the State Guest House, Akasaka Palace, when President Trump visited Japan, as well as photographs of Secretary Bessent, and these are displayed alongside photographs from the G7. In that context, we also discussed how the U.S. government views this issue. I imagine there are various discussions taking place within the U.S. government as well, and OpenAI does not necessarily know the full picture. Still, I felt that we share a common understanding in the sense that such technology must not be used by dangerous actors, and that, as democratic countries, we should cooperate in that regard. We also discussed the idea of AI defending against AI. If defense is no longer performed by white-hat hackers, but rather AI itself creates something equivalent to a patch and applies it—in other words, if AI itself carries out the process to that extent—the question becomes how the cost of that work should be treated. Is it a labor cost? It would not make much sense to pay money to AI. So what should we do? The costs involved may include electricity costs, the cost of various components for the highest-level computers, in which Japan also has strengths, and the cost of chips, among other things. We discussed such issues, as well as whether this would actually reduce headcount or the number of jobs, or whether, conversely, in some areas the number would increase. OpenAI indicated that there would also be areas in which the number of jobs would increase. We also exchanged views on whether global dominance in AI would become concentrated in particular hands, or whether the landscape would become more decentralized. Although none of this has been fully settled, we covered a range of possibilities, including whether different parties would move into areas where they have particular strengths.

Q.

I have one more question. Anthropic, rather than OpenAI, has announced that it will make technology comparable to Claude Mythos generally available. Could you tell us what the Japanese government understands about this matter, and how you view the concerns regarding making it generally available, given that such a release was originally withheld because of the risks involved?

A.

I have not confirmed that myself. However, when we asked OpenAI about reports to that effect, they said that it was unlikely that the exact same model would be made fully open. They also said that, as in this case, while OpenAI, too, is expanding access to various trusted customers, such a process would still be necessary. In other words, making a model completely open would mean making it available to anyone, and they indicated that this would probably not be the case. That said, since representatives involved with Mythos are also here, I intend to confirm this with them as well.

Q.

Going forward, does the Japanese government plan to confirm this with Anthropic?

A.

Yes. We have not received any such notification. The Japanese and U.S. governments are also in frequent contact, so we will confirm it through those channels as well. Anthropic also has representatives based here, so I think we will confirm it with them too. We have not done so yet.