Twenty-Six Thousand Rented Servers Went Quiet in One Weekend, and the Attackers' Reach Ran to More Than Seventy Brokerages

For a few days in late October, twenty-six thousand seven hundred and thirty-nine rented Windows servers stopped doing the one thing they were paid to do.

They belonged to Onamae.com Desktop Cloud, a GMO Internet product that leases virtual private servers by the month to people who need a machine that never sleeps. Retail currency traders form the natural market: the foreign-exchange market runs around the clock five days a week, humans do not, and software such as MetaTrader 4 and MetaTrader 5 will keep placing orders all night on its owner’s behalf. Doing that requires the program to store the customer’s brokerage credentials on the rented machine — login ID, password, frequently a PIN — and to read them again at every reconnect.

Break into the host, and those credentials become readable at scale. The accounts under attack were not GMO’s; they belonged to customers and were held at dozens of separate brokers. An intruder sitting inside the host could therefore reach past GMO’s own walls into institutions GMO had no authority over.

▸ The figure GMO eventually published — 26,739 servers potentially exposed — counts machines that might have been touched, not confirmed compromises. The company says as much.
▸ Fifty million dollars was the opening demand, in crypto; once that deadline lapsed, the attackers dropped to one million, offered by the twenty-seventh in exchange for sparing users a ransomware wave. Nothing has been paid.
▸ Within a week, five licensed Japanese brokers had confirmed unauthorised trades, several more had issued warnings, the police had the case, and GMO had asked the government to invoke a cyber defence law barely three weeks old.

What the episode exposed had less to do with any single firm’s competence than with the stack that modern retail trading rests on, and with the seams between the companies that own its parts.

GMO is not a small player that can be written off as unlucky. Founded in 1991 and taken online four years later under a slogan about the internet being for everyone, the group now spans roughly a hundred and thirty companies, a dozen of them separately listed, and employs close to seven thousand nine hundred people on a consolidated basis. Revenue for the 2025 fiscal year reached about two hundred and eighty-five billion yen, with group assets just past two trillion. Japan’s largest domain registrar sits inside it, along with cloud and hosting brands, an internet bank, a crypto exchange, a cybersecurity arm, and GMO Click Securities, which has repeatedly ranked as the world’s highest-volume FX broker. When an organisation of that size loses control of a product it rents to traders, the damage does not politely stay inside the organisation.

The service at the centre of the incident is, by contrast, obscure enough that most of its users probably stopped thinking about it years ago. Onamae.com Desktop Cloud is a Windows remote-desktop virtual private server — a rented box in a data centre, paid for by the month, kept switched on so that trading software can click away through the night while its owner sleeps. Nothing about the product is glamorous, and that is the point. It sits beneath the interesting layers, below the apps and the brands, and its entire selling proposition is that you forget it exists.

Commodity infrastructure earns its money by being invisible. When it works, nobody notices; when it fails, everything stacked on top of it falls at once, and the world discovers that a thing it treated as a utility was in fact load-bearing. A man in Osaka who day-trades the yen does not own a trading desk. He rents a slice of one, runs somebody else’s software on somebody else’s machine, and connects to a broker that is itself a link in a longer chain. Five reputable companies have been assembled into a personal trading operation, and the trader tells himself it is safe because each one is respectable. What he has actually built, without knowing it, is a chain — and a chain holds only until the quiet link nobody inspects gives way.

The four days that followed are worth reconstructing carefully, because the sequence matters almost as much as the outcome. So does the architecture that made a hosting invoice into a national question, the extortion that turned a technical failure into a billing dispute, and the two-year run-up during which Japan had been losing money through the same gap without quite naming it. There is also the matter of a three-week-old defence law that met its first real test, and of a chief executive who chose that moment to say in public that the internet is breaking down.

A Rented Windows Machine, a Stored Password, and the Weekend Twenty-Six Thousand Servers Went Quiet

Detection came late and, by GMO’s own account, close to accidentally. An investigation opened on the nineteenth of October after a customer asked an awkward question. Malicious programs turned up on some servers the next day, and new sign-ups were frozen. Then, at four in the morning on the twenty-first, the company began shutting the entire fleet down, plan by plan, and published its first notice that afternoon. That initial language was careful: leaked information had not, at that point, been seen being misused. A second notice, a day later, withdrew the comfort, saying that login credentials for trading accounts might have been obtained and used without authorisation.

Traders noticed before the paperwork caught up. One customer, interviewed by the broadcaster TBS, watched his balance hit zero on the twenty-first; his trade history recorded losses above four million yen inside two seconds. Roughly two hours later, an email arrived from a sender calling itself “HackingTeam,” explaining that the account had been emptied because GMO had refused to negotiate. The group’s own claim page asserted that it had destroyed accounts worth about 7.7 billion yen in twenty-four hours — an attacker’s figure, not an audited one.

Switching the servers off did not stop the losses. Stolen credentials still worked from anywhere, because the accounts sat at the brokers. GMO ended up asking customers to terminate every active session, warning that a password change alone would not help while an open session could keep placing orders.

Japan’s retail margin-FX market ranks among the largest in the world, and an unusual share of its order flow is placed by software rather than by hand. Traders in the country have long been comfortable with automated strategies, partly because the yen’s volatility rewards speed and partly because the market never closes for the weekend the way a domestic equity exchange does. A hosting product tied to trading bots therefore becomes, in practice, a single point of failure for a great many accounts at once, and the customers who rented it had no reason to suspect that their safety depended on a company they had never heard of.

The brokers, meanwhile, discovered the problem from the other direction. Five licensed Japanese firms eventually confirmed unauthorised trades on client accounts: AvaTrade Japan, Gaitame Finest and Dukascopy Japan early in the week, then OANDA Securities and FXTF on the Monday. FXTF, which runs more than 280,000 client accounts from Tokyo, said several MT4 accounts had been traded without their owners’ consent, and announced mandatory one-time passwords for MT4 and MT5 along with tighter limits on access from abroad. OANDA went further than it had days earlier, when it merely warned that logins might have been viewed; now passkeys became compulsory, and clients were told to claim directly from GMO while the broker helped them assemble trade records. Gaitame Finest had already restricted affected accounts on the twenty-first and warned that withdrawals might take longer.

Other notices split in two. Several firms found nothing on their own books — GMO Click Securities and GMO Gaika, both inside the parent group, fell into that category — while others, among them SBI FXTrade and Invast Securities, told clients to register passkeys and stop reusing passwords without confirming any trades of their own. Of eleven notices the trade press reviewed, five admitted trades had happened. The remainder, in effect, said that they might have been lucky.

Not a single broker reported a breach of its own systems, and there is little reason to doubt them. Their defences were not defeated; the customers’ rented machines were, and the failure landed exactly where nobody had agreed who owned the responsibility. A trader trusts his broker, and the broker trusts its own systems — which, this time, held up. What the broker never audited was the client’s own rented machine, because that machine belongs to the customer and the customer’s security habits are his private business. The host, for its part, trusted customers to harden their boxes and trusted itself to secure the platform beneath them. Each party performed its half of a contract nobody had signed, and the attacker simply walked through the space where those halves failed to overlap.

Regulators moved with the peculiar urgency of institutions that cannot fix anything directly. Japan’s Financial Services Agency asked FX firms, through the Financial Futures Association of Japan, to warn clients and keep a close watch on order flow for anything suspicious. That was the extent of it — watch, warn, wait. Nobody could reset the broken chain from the top, because nobody owned the whole of it.

GMO’s own admission, offered later, was that the breach had not been unforeseeable. Its in-house white-hat hackers had flagged gaps in the service beforehand, the operating team knew about them, and work was underway. Priorities got in the way, the pace was too slow, and the attackers walked into the gap that remained. That, more or less, is how most major breaches are born: not from a single catastrophic oversight, but from a known problem that never quite made it to the top of a list.

An Extortion Demand, a Refusal to Pay, and the Compensation GMO Chose to Carry

Modern extortion has drifted away from the simple lock-and-ransom model. GMO’s attackers barely used the classic kind of encryption at all. They held the customers’ own accounts as collateral and billed the company for relief.

Three demands set the terms, in sequence. Fifty million dollars in crypto assets came first, due by five in the afternoon on the twenty-first. When that lapsed unpaid, a smaller figure followed: one million dollars by noon on the twenty-seventh, with the promise that users would be spared a ransomware attack. Behind both stood a standing threat to keep firing unauthorised orders at more than seventy firms until somebody paid. GMO confirmed on the twenty-sixth that it had complied with none of it.

Crypto was not decorative. Dollars settled in digital assets move within minutes at any hour and land in a wallet with no name attached, sidestepping the banking hours and correspondent chains that make cross-border payment slow. The instrument matched the crime: fast, borderless, and just opaque enough to leave a victim guessing whether compliance was even possible before a deadline.

Three separate punishments hung over the whole negotiation. More fake orders came first, placed indefinitely at firms in Japan and abroad. Then the publication of data scraped off customers’ own PCs, bundled with a promise of ransomware. Finally a dump of internal detail about how GMO’s systems and operations actually work. Each threat was calibrated to a different audience — the traders, the company, and the company’s reputation — which is why GMO could not simply wait the attackers out.

GMO’s public answer was unusually blunt. It published the demands, confirmed it had paid nothing, and referred the matter to police with the communication logs attached. Then came the commitment that carried more weight than any statement: full compensation for realised trading losses traced to the intrusion, including where a broker refused to refund its own client, plus fees. Market losses and lost profit were excluded, as they tend to be. Every current user of the service would also receive a flat five thousand yen, roughly thirty-two dollars and frankly symbolic.

Claim forms were promised by the sixth of November. Affected users would be offered six months of free hosting on a replacement environment running on a different cloud, and the original servers would never restart. Rebuilding the underlying architecture with outside specialists was promised as well. That last decision is the most informative: when a company abandons hardware instead of cleaning it, the compromise is assumed to run deeper than any patch can reach.

The arithmetic deserves a moment. GMO can absorb the bill; a smaller host could not have. Firms carrying the most critical weight are rarely the ones with the deepest pockets, and the deep-pocketed usually got that way by keeping margins thin. Attackers exploit the gap, which is why extortion increasingly lands on the weakest link rather than the largest target. An attacker does not need to topple a giant. He needs only to find the quiet supplier the giant depends on and squeeze until the giant’s customers start shouting, at which point the giant will usually pay — not to save itself, but to make the shouting stop.

There is a second economic layer that rarely gets discussed. Compensation of this kind is not charity; it is damage control priced against the cost of litigation, regulatory attention and lost trust. By promising to cover realised losses even where a broker declines to refund, GMO effectively underwrote the weakest part of the chain — the customer — and in doing so removed the incentive for anyone downstream to argue about fault. Whether that sets a precedent other hosts will follow is an open question. Whether they can afford to is a harder one.

The Long Run-Up: Phishing, Penny Stocks, and Two Years of Damage in Japanese Brokerage

October did not invent the problem.

Japan had been losing money through the same gap for well over a year, and the only genuinely new element was the entry point. Early in 2025, account holders at several large brokers began finding their shares sold and replaced by unfamiliar names. Login IDs, passwords and PINs had been harvested through phishing and malware; the attackers logged in as their victims, liquidated the holdings, and — the distinctive move — spent the proceeds on thinly traded penny stocks. Those shares turned over in tiny volumes, so sudden buying lifted their prices, and the attackers, having accumulated the same names in advance, sold into the spike. The victim’s account did the pumping. The attacker’s wallet collected.

The technique had a logic that older fraud lacked. Japanese brokerage accounts had long been considered safe, because even a successful intruder would still have needed a bank account in the victim’s name to withdraw the proceeds. Stealing shares was one thing; getting the money out was another. The 2025 generation of attackers bypassed the problem entirely by never withdrawing anything. They sold, they bought, they sold again, and the profit stayed inside the market, arriving as a gain on positions the attackers already held.

Figures from Japan’s Financial Services Agency show how quickly the scheme scaled: 7,139 illegal transactions by June 2025, and unauthorised securities sales worth 571 billion yen. Fraudulent trading from hijacked accounts passed 500 billion yen across nearly six thousand incidents between March and May. Unauthorised access cases climbed from sixty-five in January to 1,420 in March to 4,852 in April. Against the size of the market these numbers were modest; against individual retirement savings they were not.

One example from the securities side illustrates the mechanism. A single illiquid stock that normally turned over tens of thousands of shares a day suddenly traded 4.46 million shares on March 26 and 3.31 million the following day, with no obvious reason for the surge. The explanation, most observers concluded, was manipulation executed through stolen accounts, and the price movement was the fingerprint.

Two forces produced the exposure. Deregulation in the late nineties, the so-called Financial Big Bang, pushed the country into online brokerage and left a generation of platforms optimised for frictionless login. SoftBank and Sumitomo Bank, through mergers with American firms, launched their own brokerages in 1998 and 1999, and for a quarter of a century the industry ran without serious unlawful trading of the kind that later appeared. Meanwhile the public was only beginning to move from savings into investment, frequently older, often drawn in by tax-free NISA accounts, and seldom prepared for a world in which a password is a liability. Cards and banks had spent years hardening themselves. Brokerages, comfortable after so long without trouble, had not, which made them the obvious target.

Faking Japanese used to be difficult for foreign attackers. The language’s layered scripts gave a clumsy phishing email away in a sentence, and for years that awkwardness acted as a natural moat. Generative AI removed the barrier, and counterfeit sites and messages began arriving polished enough to fool the very institutions being impersonated. Brokers then met a trade-off that remains unresolved: each additional authentication step makes an account safer and a trade slower, and traders dislike slow. Call centres filled with customers who could no longer log in, and the industry discovered that security and convenience were pulling in opposite directions.

Compensation followed the embarrassment. Firms that had resisted liability quietly began paying victims anyway, stretching rules that never strictly required it, because the reputational cost of refusal had grown. Authentication tightened, two-factor requirements spread, and the volume of fraudulent trades fell from a peak of 2,932 in April to 2,329 in May and 783 in June. The industry concluded the worst had passed. October showed it had merely been deferred.

Broader context explains why Japan proved so attractive. Credit-card fraud reached a record 55.5 billion yen in 2024, according to the Japan Consumer Credit Association, while wire fraud aimed at online banking moved 8.69 billion yen the same year, per the National Police Agency. More than a hundred thousand stolen card records were found circulating on the dark web. Against that background, brokerage accounts — lightly defended and, until recently, largely ignored by criminals — looked like an obvious upgrade.

Even before October, the pattern had begun touching institutions that had nothing to do with trading. Daiwa Securities disclosed that information on as many as 110,000 clients might have leaked after the servers of an external vendor were compromised, a reminder that the same architecture — one supplier, many downstream victims — was already producing casualties in the securities business proper.

A New Defence Law Meets Its First Real Test, and the Government Says Little

Three weeks before the breach, on the first of October 2026, the sharpest part of Japan’s active cyber defence framework took effect. Enacted in May 2025, the law allows the police and the Self-Defence Forces to reach into servers used in an attack and neutralise them — deleting malicious programs, altering settings, cutting the attacker off. The machine survives; its usefulness to whoever hijacked it does not.

GMO asked for precisely that, calling its case the framework’s first activation. The request went beyond an ordinary criminal complaint. What GMO wanted was state-led tracking, blocking and neutralisation of the attackers’ own infrastructure: the servers issuing commands and the sites where stolen data would be posted.

A private company could make the ask because the relevant clause, added to the Police Duties Execution Act, does not require the attacker to be a foreign state. All it requires is a wrongful act carried out through information technology, and an urgent risk of serious harm to life, body or property. Ordinary people’s trading accounts satisfy that test, and GMO’s notices leaned on the phrase “citizens’ assets” with the wording clearly in mind.

The framework was never designed with a case like this in view. Its architects imagined a foreign power running a sustained campaign against the state or against critical infrastructure, using the same tools and the same servers over months. GMO’s attackers were, on the evidence available, a criminal crew raiding the customers of a single private product, which fits the template only loosely. That mismatch runs through everything the government has said since.

The government’s first public answer was cool. Chief Cabinet Secretary Minoru Kihara told reporters that neutralisation would be decided by the police or the defence ministry and Self-Defence Forces “in light of the requirements, regardless of whether a victim organisation has made a request” — a polite way of noting that the law offers companies no procedure to apply, and that the decision belongs to the state. Pressed on GMO specifically, he withheld detail, saying comment might tell the attackers what the government intended. That same day, Prime Minister Sanae Takaichi convened the first meeting of a ministerial council on cyber response and promised an emergency package within the week.

What the law actually permits is narrower than the phrase “active defence” suggests. Neutralisation means removing malicious programs from a server or changing settings so the attacker can no longer use it; the hardware is not destroyed. The government presents the power as one option among several, alongside alerts, takedowns and public attribution, and it applies the same article to the Self-Defence Forces with different wording and conditions. For ordinary people’s accounts, the police clause is the one that could plausibly apply, since the SDF provisions are drafted for attacks on government and critical computers.

Servers located abroad thicken the process considerably. A National Security Council review of response policy, consultation with the Foreign Minister, and prior approval from the independent Cyber Communications Information Oversight Commission all precede any intervention. Where speed is essential, the police can notify the commission after the fact rather than before, but the ordinary case follows the longer route. GMO describes its attackers as operating across borders, so their equipment may well sit outside Japan, which triggers the full sequence.

International law adds complications of its own. Akihiro Yamaguchi, a researcher at Japan’s National Institute for Defence Studies, has written that access and neutralisation measures can collide with the sovereignty of whichever country hosts the target, and that the two doctrines which might excuse that collision — countermeasures and necessity — carry heavy conditions. Countermeasures need a prior wrongful act by another state, which is awkward when the perpetrators are a private group; necessity, the easier route, still demands an essential interest and a grave, imminent peril. The closest precedent is American: in 2021 the FBI, on court authority, remotely deleted web shells from compromised Microsoft Exchange servers on US soil. Even that was narrower than what Tokyo is weighing.

Opinions among governments on the underlying question are unsettled. The Netherlands has argued that mere disruption is not enough and that something like a total loss of internet access or serious damage to financial markets would be required to justify necessity. Germany and the Czech Republic take a broader view, allowing for the loss of critical functions rather than only physical damage. Whether losses from unauthorised FX trades reach that threshold is exactly the kind of question Japan would have to answer for the outside world if it acted.

One further constraint deserves mention, because it limits what the law can deliver. Neutralisation cannot bring back money already lost. Deleting a malicious program or blocking a command server may stop further damage, but balances already drained by unauthorised orders stay drained, and the victims’ recovery runs through GMO’s compensation scheme rather than through any state action. GMO has asked customers to contact their brokers about cancelling trades or obtaining refunds and has said it will cover whatever remains. The government’s intervention, if it comes, is a fence, not a rescue.

As the days passed, the “urgent need” that justifies intervention also had a habit of receding. GMO had shut its servers, told customers to change passwords and terminate sessions, and begun the slow work of rebuilding on a separate cloud. Each of those steps made the attackers’ position harder and, by the same token, made a rushed state intervention less obviously necessary. The clock was working against the case for action even as the case for action was being argued.

Cheap AI, Saturation Attacks, and a Chief Executive’s Warning That the Internet Is Breaking Down

On the twenty-fifth of October, GMO’s founder and chief executive, Masatoshi Kumagai, published a long post under his own name, and in it said something chief executives seldom say in public. The internet, he wrote, is “already starting to break down.”

His argument concerns cost. Open-weight AI models, freely combined into attack tooling, have turned work that once demanded real expertise into something a competent amateur can run at volume. As evidence he cited his own group’s research: an autonomous penetration-testing tool built for defence, called ARTEX, had been turned against roughly twenty Japanese companies by attackers. He called the moment an “AI cyber-saturation attack” — defenders losing not on skill but on sheer numbers.

Saturation is the right word for it. Traditional defence assumes that attackers and defenders are roughly matched, and that skill decides the outcome. When automated tools can probe millions of targets at once, the defenders’ problem stops being intellectual and becomes arithmetic. There are not enough engineers to review every alert, not enough hours to patch every system, and not enough budget to cover every surface. Attackers, meanwhile, pay almost nothing per attempt, because the marginal cost of one more probe is effectively zero.

When a serious PHP vulnerability surfaced, exploitation attempts moved from a trickle to millions within hours of disclosure, fast enough that Japan’s prime minister and the FBI issued warnings, and Cloudflare’s chief executive described the same jump. Many small sites took themselves offline to avoid the sweep. The same week, a flaw in the way an Indian certificate authority generated its signing keys forced companies worldwide to react and cut a fifth off that firm’s share price in one session, while a cloud platform called Railway paused all customer deployments after a breach of its own.

Read together, those incidents describe a pattern rather than a series of accidents. Each involves a single component that quietly supports a great deal, each was compromised through a small gap, and each produced damage out of all proportion to the flaw. The certificate authority matters because so much of the web’s trust rests on a handful of signing keys; the cloud platform matters because thousands of businesses deploy through it. When several such components wobble in the same week, the internet looks less like a stable network and more like a set of load-bearing pillars that nobody has inspected.

GMO’s official submission put the case more bluntly: “Defense and after-the-fact response by a private company alone cannot eradicate the plundering of citizens’ assets by sophisticated cybercrime groups operating across borders.” That is the sentence which converts a hosting incident into an argument about national capacity, and it is clearly aimed at the government rather than at GMO’s customers.

Kumagai also addressed the obvious objection — that active cyber defence is a state matter and should not move because a private company asks it to. He conceded that GMO is not a conventional infrastructure operator of the kind the law had in mind, comparing it to an electricity or gas company, and acknowledged that the government decides on its own initiative. What he offered in return was material: full logs, compromise data, and a disclosure request filed through the courts to obtain the attackers’ trail.

Kumagai conceded, too, that the breach was not unforeseeable. GMO’s own white-hat hackers had flagged gaps in the service beforehand; the operating team knew; work was underway, but too slowly against competing priorities. He closed by tying the episode to the new defence law and to public-private cooperation, and by warning ordinary citizens that the next emptied account could easily be theirs.

There is a wider question buried in his post, and it is the one that will outlast the incident. If attacking has become cheap and defending has stayed expensive, then the balance between the two is shifting in a direction that favours the attacker, and no single company, however large, can correct it alone. The alternative — a state that reaches into hostile servers and neutralises them — carries its own risks, from sovereignty disputes to the simple possibility of doing harm to a machine that turns out to be innocent. Japan has now been handed a live test of which cost it prefers.

What the Week Left Unanswered

Several things remain unresolved, and they are worth separating from what is known.

Whether a single crew sits behind the whole campaign is unclear. GMO describes “sophisticated cybercrime groups operating across borders” in the plural; the attackers’ own page names only “HackingTeam.” The 7.7 billion yen figure comes from the attackers, and the reach of more than seventy firms comes from GMO; neither has been independently audited.

How many servers were genuinely compromised, as opposed to merely exposed, is unknown, and GMO says so itself. No broker has published what was lost on its own book. Whether Japan will neutralise servers abroad, and how quickly, remains open; the government has committed to nothing specific.

More confidence is possible about the structure. A commodity hosting product had gathered tens of thousands of brokerage credentials in one place, sat low enough in the stack that few people watched it, and connected to institutions that had no responsibility for it. None of that is unique to GMO, and none of it has changed since October.