On 9 October 2026, NHK ran a feature on the data leaks landing in Japan almost every day. The caption read 「止まらぬ漏えい」 which means the leaks that will not stop. The number next to it was up to 27 million personal records, disclosed by companies on that day alone.

The night before, I was at dinner with a group where the same topic came up. One of the people at the table was a Times Car member. His driving licence data was in the leak.

Watching that feature, I thought about it again. AI has lowered the cost of launching an attack to the point where it costs almost nothing to write a script and point it at a target. It has also lowered the payout an attacker needs before an attack is worth the effort. Defence has moved the other way. Every new requirement costs money, people, and time, and the bill keeps growing.

What NHK Reported on 9 October

These are the main disclosures published that day:

In the first half of the month there were also Yakiniku King (about 10.78 million records), Daiichikosho (about 8.72 million), and Lawson (more than 2.15 million).

The JR East and View Card leaks did not come from their own systems. IDC Frontier, a SoftBank subsidiary, was hit by ransomware on 7 October and 495 companies and local governments were affected. IDC Frontier says that some of the customer data will be difficult to recover.

The National Police Agency (NPA) runs internet sensors that record suspicious access. Its latest cyberthreat report says that in the first half of 2026, each sensor IP address recorded 13,687 suspicious accesses per day. That figure had stayed between 7,000 and 9,000 since 2021, and it has risen sharply. The same report counts 123 ransomware cases in the half year, the most on record.

An expert interviewed on the programme said attackers have used AI to automate the work and now aim at many targets with short gaps between attacks. He also said that getting leaks down to zero is, unfortunately, no longer realistic.

Someone at Our Dinner Table Was Affected

Times Car detected unauthorised access on 25 September. The leak covered about 6.6 million accounts. It included people who had already left the service and people who had applied but never finished signing up. About 1.6 million of those accounts included images of identity documents such as driving licences.

A leaked email address is a nuisance. A leaked photo of your driving licence, next to your name, address, phone number, and birth date, is a kit for opening a bank account or applying for a loan in your name. That is why people rushed to file declarations with the credit bureaus CIC and JICC, at a cost of about ¥2,000 (here is how the filing and the fees work).

At dinner, this was no longer a news item. It was a person at the table who now has to watch his credit record for years because he once rented a car by the hour. The company pays the investigation costs. He pays with his time and his fees, for something he did not do.

At that same dinner I said this. No company can honestly say "we will never leak" any more. The honest stance is "we will leak at some point, so what do we do then". This is not new. If you have worked in IT and with the internet over the past decade, it should already be your default assumption.

Why Attacks Got Cheap

I wrote last December about vibe coding for small businesses. My point was that a three-person team could start automating its work for under ¥10,000 in the first month. The same arithmetic applies to attackers.

Skill is no longer the barrier. In August 2025 Anthropic published a threat report on how criminals used its own model, Claude. One UK-based actor developed and sold ransomware kits for $400 to $1,200 each. According to Anthropic, he did not appear able to implement the encryption himself.

Scale is no longer the barrier either. In a November 2025 report, Anthropic describes a state-sponsored campaign in which the AI performed 80 to 90 percent of the work, with humans stepping in at perhaps four to six decision points per attack. This matches what the NHK expert described: many targets at once, short gaps between attacks.

The payout bar has dropped. When an attack costs a skilled team several months, it only makes sense against a target worth millions. When it costs a script and an afternoon, a much smaller haul is enough. GMO Research & AI found that someone had exchanged members' points on its infoQ survey site for about ¥2.87 million in Amazon gift codes. The targets are no longer only banks and governments. They are a karaoke chain, a used-book shop, and a sushi chain's app.

Anthropic wrote the same thing in its announcement of the Anthropic Cyber Mission on 8 October: "The cost of exploiting vulnerabilities has dropped, while verifying, disclosing, and fixing them is slow and still depends on people".

Why Defence Got Expensive

On 9 October the government asked businesses to introduce multi-factor authentication (a second check on top of the ID and password) and to apply security patches, among other measures. IPA, Japan's information security agency, published its checklist items the same day. Every measure is sound. Every measure needs money and people.

The cost after a breach is worse. In the NPA's survey of ransomware victims, 21 of the 35 that answered spent ¥10 million or more on investigation and recovery. Of 40 that answered, 29 could not restore their systems from backup.

The supply chain is a problem too. Daiwa Securities and Citizen were affected through their vendor, Scala Communications. JR East and View Card were affected through IDC Frontier. You can do everything right inside your own walls and still lose your customers' data through a vendor's server.

Then there are the costs that land on individuals. The Financial Services Agency asked banks and other financial institutions to tighten identity checks. From April 2027, the law already ends identity checks based on uploaded photos of ID documents. Checks will move to reading the IC chip in the card instead. The FSA asked institutions not to wait for that date. It is the right move, and it means system changes at every institution. Meanwhile the person whose licence image leaked pays the CIC and JICC fees himself.

The attacker's cost goes down with every model release. The defender's cost goes up with every new requirement, every vendor, and every incident. That is the imbalance I mean.

AI on the Defender's Side

It is not only bad news. On 8 October (US time), Anthropic launched a cyber defence programme for critical infrastructure. The programme gives its latest Claude models, on-site engineers, and threat research to the providers that protect power grids, water systems, and transport. Anthropic's announcement lists 11 founding partners, including Hitachi, CrowdStrike, Palo Alto Networks, Accenture, and Rockwell Automation. Hitachi says in its own press release that it will contribute its knowledge of operational technology (OT) from the railway and energy sectors.

This connects to something I wrote in July about the OpenAI and Hugging Face incident. After that attack, Hugging Face tried to use commercial AI models to analyse its logs. The models refused, because the logs looked like attack code. The attacker had no such restriction. I called that an asymmetry: only the defenders had to ask for permission and were told no. Anthropic has since widened its scheme for giving vetted defenders access to its most capable models.

Anthropic also makes a forecast I find honest. It expects AI to favour defence in about two years. In the near term, it says, that may not be true.

My concern is the near term. Only large companies can join the programme. Hitachi will be fine. But most of the companies on NHK's list are consumer businesses, and many smaller businesses sit behind them. In my vibe coding article I imagined a 15-person wholesaler in Osaka with ¥50,000 a month to spend on AI. That company has no security team. It will not be in a frontier defence programme. And the attacks are now cheap enough to reach it.

The Real Fix: Design Data That Cannot Be Used Against You

All of the above is about defence. The deeper question is about the data itself, and there are two parts to it.

The first is how the data works once it is out. At that dinner I also said this: if a leak is a matter of when, the data we hold should be useless to a criminal on its own. A leaked photo of a driving licence should not be enough to open a bank account or take out a loan in someone else's name. That second-order damage is a design failure, not only a security failure.

Estonia's e-Residency card shows what that design looks like. The official explanation says the card does not even display a photo. The card number on its own does nothing. To log in to a service or sign a document, you need the physical card in a reader and a PIN: PIN1 to prove who you are, PIN2 to sign. Estonia states on its page about the security codes that the card cannot be used without these codes. If a picture of the card ends up on the internet, nothing much happens. That is the standard we should aim for.

On the day I wrote this article, a news story appeared saying that identity checks are moving from images to the IC chip.

The second part is how much data we collect in the first place. I learned this at Xoxzo, the cloud telephony company I ran. We took the time and effort to get certified for the Privacy Mark, Japan's certification for companies that handle personal information to the JIS Q 15001 standard, and we kept renewing it.

The first thing the process made us do was a full inventory of every kind of customer data we took. For each item, we had to justify the business reason for holding it. The more sensitive the data, the more work it took to protect and document it. We were a small company without much in the way of resources, so we had to be creative. We learned to use less customer data and still do more for the customer.

I carried that habit into PuchiDen. The service asks for an email address and nothing else, because making a phone call should not need more than that. Kafkai and KaiMail follow the same rule: take the minimum from the customer.

The problem is that most businesses are rewarded for doing the opposite. The more they know about a customer, the more they can cross-sell, and the more they can offer partners in co-marketing deals. So consumers hand their details to dozens of services, many of which they barely use. Every one of those copies is one more place for the data to leak from.

What We Should Actually Do Now

If the cost of defence keeps rising, start with the measures that cost nothing and still work. The point is not to buy something new, but to reduce the data you already hold.

For Businesses, Especially Smaller Ones

  1. Take stock of every piece of customer data you hold. Data you do not hold cannot leak. List every field you collect and write down the business reason for each one. This is the same exercise the Privacy Mark made us do at Xoxzo, and you do not need the certificate to start it today. Times Car says the law requires it to keep transaction records for about seven years after members leave. Where you cannot justify a field, delete it and stop collecting it.
  2. Stop storing ID document images if you can. If you need to confirm identity, keep the result of the check, not the photo. IC chip reading lets you verify a card without keeping an image of it.
  3. Turn on multi-factor authentication everywhere you can. It is the cheapest control on the government's list. It costs a little friction and almost no money.
  4. List your vendors, and know what each one holds for you. You cannot audit IDC Frontier yourself. You can know which of your data sits there, and whether you have your own backup.
  5. Test the backup. 29 of the 40 that answered could not restore from backup. A backup you have never restored is only a hope.
  6. Assume a breach and plan the response. The NHK expert said it, and JR East's president said at a press conference (in Japanese) that a company must also plan for what happens once an attacker is inside. Decide now who you call, how you tell customers, and which system you switch off first.

For Individuals

  1. If your ID document image leaked, file a declaration with CIC and JICC (an English guide to filing with both, and the fees). An ID document with your photo can be used to apply for a card or a loan in your name. A declaration leaves that fact on your credit record. Lenders check it, so they will verify your identity more carefully than usual. You pay a fee, but that is cheaper than someone taking out a loan in your name.
  2. Do not reuse passwords. Skyticket leaked about 4.13 million passwords. If you reused that password anywhere else, change it there too. Turn on multi-factor authentication as well.
  3. Expect better phishing. The leaked email addresses will be used. The NPA also warns that phishing is getting better, and experts say AI is removing the barriers for would-be scammers, so bad Japanese is no longer a reliable warning sign. I wrote about this from the receiving end when someone spoofed the kafkai.com domain in a phishing campaign. The rule is the same: do not click links in unexpected emails. Open the official site or app yourself.
  4. Sign up for less. A traveller NHK interviewed at Haneda had registered with Skyticket for the first time, for a business trip. His data leaked shortly afterwards. He said that registering for too many things is a problem. I agree.

A last note about my own position. I run small services myself: Kafkai, KaiMail, and PuchiDen. I am the small company in this story, not Hitachi, so this article is also partly a checklist for myself. The economics will not reverse because we wish they would. What we control is how much we hold, how quickly we notice, and how ready we are for when it leaks.