The Urgent Need for US-China Cooperation on Super-Intelligent AI Security

Under normal circumstances, I would be writing about the geopolitical implications of the Iran conflict. However, I am compelled to write this piece unexpectedly, after which I will return to my previous topic.

In this article, I want to discuss a surprising development in the field of AI, or artificial intelligence. This week, an event occurred in the AI sector that few predicted would arrive so soon.

It is impacting not just technology or the economy, but geopolitics as well.

Last Tuesday, the AI company 'Anthropic' announced the release of a new version of its AI model, called 'Claude Mythos Preview.' However, the new model was not made available to everyone.

It was provided only to a limited group of about 40 technology companies, including Google, Broadcom, Nvidia, Cisco, Palo Alto Networks, Apple, JPMorgan Chase, Amazon, and Microsoft. Some of the companies Anthropic provided its new version to are even its competitors.

The new AI can not only write software code much more easily than any currently available model, but it can also easily identify vulnerabilities in almost all popular software systems in use globally.

There is a specific reason why it provided its product to its competitors first. Anthropic's new AI model is not just a minor change; it represents a 'phase shift' or a leap forward.

The new model has shown that it could have significant positive and negative impacts on cybersecurity as well as America's national security.

While developing its new model 'Claude Mythos,' Anthropic discovered something surprising. Its new AI can not only write software code much more easily than any currently available model, but it can also easily identify vulnerabilities in almost all popular software systems in use globally.

The danger now is that if this tool falls into the hands of bad actors or malicious forces, they could hack almost every major software system in the world. This could include all systems belonging to large corporate groups. Do not take this as a joke; I mean all systems.

On the eve of the new announcement, representatives from leading US tech companies were in discussions with the Trump administration. The discussion covered the potential impact of Anthropic's new AI version on software systems in the US and other countries, according to someone who participated in the discussion.

As Anthropic stated in its written statement on Tuesday, the Mythos Preview had already identified many high-severity vulnerabilities last month. The systems where such vulnerabilities were found included every major 'operating system' and 'web browser.'

Anthropic's statement says, 'Given the pace of progress currently seen in AI, it seems it will not take long for such capabilities to spread further. There is a high probability that such developments in AI will not remain limited to the hands of individuals or groups operating it safely. Consequently, there could be serious impacts on the economy, public safety, and national security.'

The language of this Anthropic statement is not simple or straightforward. In my understanding, Anthropic has signaled that super-intelligent AI is arriving faster than expected.

Anthropic has named one of its groups 'Project Glasswing.' That group works with large and trusted technology companies and organizations working on sensitive infrastructure, such as banks. It helps strengthen the security systems of such organizations. It also helps leading tech companies identify various vulnerabilities and problems within their systems and addresses those vulnerabilities.

Regarding the latest context, Anthropic said, 'We do not have plans to make Claude Mythos Preview generally available. However, our goal is to enable our users to operate models in the Mythos class very safely. It is also to prepare for the countless benefits that other highly capable models might bring, not just in cybersecurity.'

The language of this Anthropic statement is not simple or straightforward. In my understanding, Anthropic has signaled that super-intelligent AI is arriving faster than expected. We already knew that writing software code was becoming surprisingly easy even for someone with basic computer literacy.

But we did not even anticipate that Anthropic's new model would be able to find errors in existing code and discover ways to exploit them so quickly.

According to Anthropic, its new model has found serious risks in every major 'operating system' and 'web browser.' Many of these systems run the world's electrical transmission systems (power grids), water supply centers, airline reservation systems, retail networks, various military systems, and hospitals. (To put it simply, by using the Anthropic model's AI, even an ordinary person would have the capacity to affect, hack, and destroy any sensitive system in the world.)

If the new AI tool Anthropic has prepared were to become widely available for everyone to access, the capacity to hack any major infrastructure system would increase suddenly. The capability previously held only by a few experts and intelligence agencies would reach everyone's hands.

Criminals, terrorists, and rogue states would easily gain the ability to 'hack' sensitive systems.

It is because of this risk that Anthropic has provided only carefully controlled versions to its specific software providers.

I am not blowing this issue out of proportion or creating a scare. I am writing with genuine seriousness. In the coming days, small children could use such AI. Parents must also be prepared for that from now on.

Imagine a scene!

You ask your child, 'Son/Daughter, what did you do today after school finished?'

In response, your child replies— 'Mom/Dad! My friends and I shut down the power grid today. But I am very hungry, what have you made for a snack today?'

It is because of this risk that Anthropic has provided only carefully controlled versions to its specific software providers. It has done so with the aim that they can identify and fix vulnerabilities in their systems before people with bad intentions or your children do.

In such a serious situation, I would like to discuss, understand, and learn from my technology mentor, Craig Mundie. He is a former director at Microsoft and a member of former US President Barack Obama's Council of Advisors on Science and Technology. He has also written a book about AI titled 'Genesis' with Henry Kissinger and Eric Schmidt.

We discussed the latest problem. In our view, no country in the world can solve this problem alone anymore. The solution option might surprise you.

Yes, such a problem now needs to be solved by the two AI superpowers working together. Those two superpowers are—the United States and China.

It has become essential for them to cooperate. Otherwise, bad actors in world politics and society could gain this level of capability. By then, it will be too late.

Now, such powerful tools appearing in the course of AI development put the systems of both countries at risk. It makes them increasingly unsafe against both national and international terrorist groups and other adversaries.

In the coming days, new AI tools can easily become a major threat. That threat is certain to be much larger than the mutual threat between these two countries.

This is not a normal situation. Just as the 'Mutual Assured Destruction' and the need for nuclear non-proliferation were required during the nuclear arms race, today's time has become at least as important.

This is a bigger issue than Russia's concern right now. The situation has arrived where the US and China must work together to prevent humans or autonomous AIs from misusing such terrifying technology.

This is a very important and urgent context. It should be included as a main agenda item in the summit between US President Donald Trump and Chinese President Xi Jinping in Beijing next month.

First, the release of such super-intelligent models must be carefully controlled.

Mundie's view was, 'Previously, such matters were dominated by large countries, militaries, large companies, and large criminal organizations that could spend a lot. Now, such capabilities to develop sophisticated cyber hacking operations can be easily available to small players or ordinary individuals. Cyber attack capability is also becoming democratized.'

According to him, responsible governments now need to do three things mandatorily in collaboration with the companies building such AI tools and software infrastructure:

First, the release of such super-intelligent models must be carefully controlled. Also, it must be ensured that such models reach only responsible governments and companies.

Second, it is necessary to distribute defensive tools to avoid risks as soon as new models arrive. By doing so, all vulnerabilities in the software running their critical infrastructure can be identified and managed properly before hackers get them.

Third, we must work with other responsible countries, including China, to build secure and protected working circles, both public and private, within all special networks. By doing so, companies and governments will be able to migrate all their important services there and be safe from future hacking attacks.

We have arrived at a remarkable moment in history. Two major events occurred on April 7. The US announced the suspension of bombing in Iran. Anthropic's 'Claude Mythos Preview' model has been released by its technical partners for a limited number of companies with caution.

Let's see which event history will remember more! It will certainly be interesting to see in the future.

(From The New York Times)