The GMO Breach and Security in the Age of AI — Thoughts from a Security Specialist

Hi, Kei here.

My first post in a while is on a slightly heavier topic: the unauthorized access to GMO Internet’s “Onamae.com Desktop Cloud” service in October 2026. GMO Internet is one of Japan’s largest internet infrastructure companies, and Onamae.com is one of the most widely used domain registrars in the country.

News about personal data leaks had been coming one after another lately. So when I first saw this story, on Yahoo! News or NHK, I took it as “another one of those.”

My view changed when the attackers started issuing threats with deadlines. The other thing was who had been hit: GMO, a company whose business is providing internet services. I manage this blog’s domain, keikun.info, through Onamae.com myself. This was not something I could watch from a distance.

The more I looked into it, the more it felt like a turning point for computer security. As a Registered Information Security Specialist (a Japanese national certification), and as someone who has run a home server for 23 years, I want to write down what I think.

What happened (as of November 1, 2026)

First, the facts. Based on GMO Internet Group’s announcements (in Japanese) and news reports, here are the key points.

October 19: GMO began investigating after a customer inquiry. On the 20th, it found malicious programs installed on some servers
October 21: GMO shut down all plans of “Onamae.com Desktop Cloud,” a cloud service used for automated FX (foreign exchange) trading. Up to 26,739 servers may have been affected
October 24: GMO announced that customer data had been taken out. This changed its earlier statement that no leak had been confirmed
The people believed to be the attackers demanded about 8 billion yen (US$50 million). GMO refused
The attackers set a deadline of noon on October 27, threatening to publish data and launch ransomware attacks on users’ PCs. That day, data said to belong to one customer was published (no ransomware attack has been confirmed)
October 28: the domain of the site used to publish the data was shut down. GMO set up a bounty of up to US$1 million in total for information leading to the culprits

On the users’ side, there has also been damage from unauthorized trades in FX accounts (in Japanese). GMO has announced a 5,000 yen (about US$33) payment to every user, plus compensation for losses from unauthorized trades.

GMO also asked the Japanese government to invoke “active cyber defense” (in Japanese) under a new law that took effect on October 1. The Tokyo Metropolitan Police are investigating on suspicion of violating the Unauthorized Computer Access Act, but as of October 31, no suspect had been identified or arrested. GMO’s CEO Masatoshi Kumagai and others are scheduled to hold a press conference on November 2.

Then on November 1, another GMO subsidiary announced that, due to a configuration mistake, personal data of about 45,000 people had been viewable from outside for nine years. That one was not unauthorized access.

This is an ongoing story, so some of what I write here may change as new reports come in.

What made this different from the usual leak

There were three main reasons this case looked different from the leak news before it.

The first was how theatrical it was, starting with threats. Demand a ransom with a deadline, and publish data every day at noon if it isn’t paid. Attackers setting an exact time and putting on a show for the public felt completely different from the usual “a leak has been discovered” kind of report.

The second was that the target was GMO. It provides the basic building blocks of the internet, such as servers and domains, and I think of it as one of the experts of the internet world. The group even includes a white-hat hacker company, “GMO Cybersecurity by IERAE.” An expert company like that getting hit is a serious fact.

The third was that the weakness had been pointed out in advance. In a blog post on October 25, Kumagai apologized, saying that the company’s in-house white-hat hackers had flagged issues beforehand, and that while countermeasures were underway, they were not deep or fast enough (in Japanese). Then on October 31, NHK and Nikkei reported that the group’s security subsidiary had pointed out a vulnerability in the system as early as April 2024, but it had not been properly addressed.

In other words, the experts noticed the hole two years ago and told the company. Even so, it was not properly dealt with, and then this incident happened. It has not yet been announced whether the entry point in this attack was the same one flagged two years ago. The November 2 press conference may explain that.

Following the news, it could look like a battle: “a shadowy hacker group vs. GMO’s white-hat hackers.” But after learning this third fact, I saw it differently. It was not that they lost on technical skill. The white-hat hackers’ voices were not acted on inside the company. That is how it looks to me.

Why were the experts’ warnings left unused for two years?

So why was a warning from two years ago not acted on? The details have not been made public, so what follows is my own view.

I think it comes down to the internal ties and pressures inside a company.

Say an expert points out, “There’s a hole here.” Closing it takes people, time, and money. In some cases you may need to stop a service or rebuild part of the system. Yet after all that, sales don’t go up by a single yen. The only thing you get from closing the hole is a state where “nothing happens.”

Meanwhile, from a management point of view, numbers like revenue, profit, and operating margin inevitably come first. Launching new services, gaining customers, improving margins. When work tied to those numbers competes with security warnings for the same budget and people, it is easy to guess which one wins. Money spent on security gets called a “cost,” and it ends up ranked behind revenue and profit. On top of that, the fact that “nothing has happened so far” makes it easier to justify putting it off.

This isn’t just about GMO. After about 20 years working in the IT industry, it is a pattern I see often. Experts’ opinions tend to come second in the eyes of management. The people raising the issue are right, but because they cannot show that in numbers, their turn rarely comes.

GMO has a white-hat hacker company inside its group and sells security as a business to others. Even that company could not make full use of the warnings for its own service. Having experts on staff and reflecting their voices in management decisions are two different things, I think.

And I feel the price of this “putting it off” has grown much larger than before, because the attackers’ tools have changed.

AI built for defense is being used for attacks

Around the same time as this incident, there was an announcement that showed that change clearly.

On October 19, GMO Cybersecurity by IERAE announced that after investigating the infrastructure behind attacks on about 20 Japanese companies, it had found that “ARTEX,” a tool that uses AI to run penetration tests autonomously, was being used (in Japanese). Eight of the targeted companies had announced data leaks in September and October 2026 (in Japanese). That matches my feeling that “there sure has been a lot of leak news lately.”

ARTEX was originally an open-source tool published on GitHub by a Chinese-speaking developer for learning and research. The author has since taken it down, but modified versions are reportedly circulating (in Japanese). A tool made to find weak spots so they can be defended is being used as a tool for attack.

Until now, attackers needed a fair amount of skill and time to find a hole. A hole that had been put off might stay unnoticed for a while. But if AI can run penetration tests automatically, the effort needed to find holes drops sharply. The chance of a hole being found while it is being put off must be much higher than before.

As it happens, October 19, the day IERAE announced this analysis, was also the day GMO began investigating the Desktop Cloud issue. No connection between the two has been announced. Still, on the very day the group’s white-hat hackers published their analysis of attacks on others, an investigation was starting on one of the group’s own services.

Kumagai has also warned in his blog that open-weight AI models are being built into attack tools and spreading. Whether AI was used in the attack on Desktop Cloud has not been announced. Even so, my guess is that this was an AI-assisted attack too.

AI is like a chef’s knife

Put that way, AI might sound like the villain. I don’t see it like that. AI is just a tool. Like a chef’s knife, whether it does good or harm depends on the person holding it.

I also work on security for my home server by consulting with Claude Code. As I wrote in an earlier post, I had AI review my router config, and it found mismatches in my settings that I had missed. Ask it to “find the cause,” and AI will list one candidate after another.

But in the end, a human makes the call. If that human lacks expert knowledge, they cannot make the right decision, because they cannot push back when the AI suggests something wrong. If you can’t push back, all you can do is believe it.

People also make decisions using information they never typed into the AI. We can read the room. If you fed every piece of information into the AI, its accuracy might improve to some extent, but in reality that rarely happens. I still think humans have the better overall judgment.

Lately I see more examples of superficial security assessments done with AI. An assessment that swallows everything the AI says will hit a wall sooner or later. Humans should hold the basic knowledge, and use AI efficiently to make the most of that knowledge. That is the right way, in my view.

Back to GMO. GMO had people with the knowledge. The fact that they could point out the hole two years ago shows there were experts with solid fundamentals. What was missing was not knowledge, but a system for listening to them and acting. In the age of AI, what we need is both: people with knowledge, and an organization that actually listens to them.

“Defenders are at a disadvantage” belongs to the pre-AI era

So can people with knowledge, using AI, stand up to attackers who also use AI? I believe they can.

There is a saying in security: “An attacker only needs to find one hole, but defenders have to close every hole.” It has long been obvious that defending takes more advanced skill than attacking. The news made a big deal of the AI threat this time, but attacking being easier than defending is nothing new.

But I think that saying belonged to the era before AI. Finding and closing holes is exactly what AI is good at. If attackers use AI, defenders who use AI should be able to protect themselves far more efficiently than before.

Between AIs with the same knowledge, attack and defense become an even match. Then what decides the outcome is the skill of the person giving instructions behind the AI. And no matter how quickly that person finds a hole, nothing is protected if no one decides to close it. Skill, and the decision to use it. This case seems to show that you need both before you can defend anything.

20 years ago, I was chasing attackers on my home server

How has the relationship between attack and defense changed over the past 20 years? Let me look back through the history of my own home server.

In 2005, I wrote a post on this blog called “Unauthorized Access.” After a server I managed for my university seminar was hit by SSH password attacks, I restricted SSH on my home server to connections from inside Japan (.jp domains). Even so, I was attacked openly from a Japanese IP address. Reading the logs, I looked up the attacker’s provider by hand with nslookup and whois. Back then, noticing an attack and investigating the attacker were all done by hand.

Twenty years later, attacks still reach my home server. But hardly anyone would deliberately target a niche server with as little traffic as mine out of interest. What comes is probably automated bots. What has changed is how I defend it. These days I rely on AI’s help more and more.

There is also something I have always believed hasn’t changed: a serious attack doesn’t come out of nowhere. First, attackers comb through every hole. Then they go quiet, wait for a while, pick their moment, and strike all at once. In a sense, it is like war. The ones pulling the strings are people. AI is still just a tool.

Looking at this case that way, the weight of those two years stands out. We don’t yet know whether the entry point in this attack was the same one flagged two years ago. But if a hole really stayed open for that long, the attackers had plenty of time to probe, wait, and pick their moment. Putting something off also means giving the other side time.

Security is like a home alarm system

Thinking it through this far, it all comes back to how we think about money. In the end, the reason a warning from two years ago was put off is a question of how we value spending on security. And this is what I most want to get across in this post.

I think of security as something like a home alarm system. Install an alarm system, and if no burglar ever comes, you just keep paying the security company and gain nothing. And if a burglar does come and you stop them from stealing 1 million yen, your loss is zero, but you haven’t made any profit.

It is the same with computers. If data is stolen and you suffer damage, that is simply a loss. If nothing happens, then nothing happens. Security measures never produce a profit, no matter how far you take them.

I feel a lot of people don’t understand this. Quite a few think the right amount to spend on security is zero. That is exactly why experts’ warnings get labeled a “cost” and put off.

And when something does happen, the company is not the only one who pays the price. Speaking to NHK, a man in his 40s who lost about 40 million yen (roughly US$270,000) through trades he never made said he has been “fighting fear the whole time.” A cost that goes unpaid doesn’t simply disappear. It changes shape and lands on the users who trusted the service.

For readers who aren’t security experts

That got a bit technical. So what could users have done? I want to write something for readers who aren’t familiar with security, too.

Users had no way to prevent the GMO case

For this GMO case, I think there was basically nothing users could have done to defend themselves.

GMO has not yet announced how the attackers got in. But the attackers have named two things in their statements. One is that the remote desktop connection files GMO distributed were set up to share the drives of users’ own PCs. The other is that a shared administrator account was used to manage the servers (an article that separates the attackers’ claims from GMO’s announcements, in Japanese). GMO has confirmed neither, and neither has been verified.

If the attack did come from those two combined, both are problems with the provider’s settings and management. Users were just connecting the way they had been told to, and there was no way for them to notice and prevent it. So I basically believe this was GMO’s responsibility: it should have secured the service properly.

The wave of personal data leaks is a separate issue

On the other hand, the personal data leaks happening all over the place lately are a completely different story from this GMO case. I don’t think they should be lumped together.

In those leaks, things like the email addresses and passwords you registered with a service can get out. With these, there are things users can do to keep the damage from spreading.

Start by not reusing passwords

The first thing worth doing is using a different password for every site.

If you reuse the same password, a leak from just one site lets attackers into the other sites where you use it. If every site has a different password, a leak from one site stops at that site.

I use a different random string as the password for every site, and keep them all managed.

If you are reusing passwords right now, it may be worth reviewing them once. Even if you are registered on a lot of sites, a day or two should be enough. And once it’s done, it’s done. It doesn’t take expert knowledge, and it makes a big difference.

Summary

With leak news coming one after another, I first thought this was “just another one.” But as I looked into it, I found a hole that had been flagged two years earlier, a structure in which experts’ voices get put off in favor of the numbers, and the reality that AI tools built for defense are being used for attacks.

AI is a tool, just like a knife. In an age when attackers and defenders hold the same tools, what decides the outcome is the knowledge of the people behind them, and organizations that listen to those people and act. And security is like a home alarm system: it produces no profit. Even so, it is not something you can cut to zero.

Users could not have prevented the GMO case, but with the wave of leaks, there are things users can do. If you aren’t familiar with security, starting by reviewing reused passwords is plenty.

I passed Japan’s Information Security Specialist Examination in 2011, registered as a Registered Information Security Specialist around 2018, and still keep my registration up to date. It is a certification that demands advanced knowledge, but there is a gap between the value of that knowledge and how often it gets recognized at work, and it often makes things hard.

Even so, I write posts like this hoping the day will come when this way of thinking gets the attention it deserves. I hope this case helps more people realize it soon.

— Kei