inthesingularity.ai
Calculator

C.005The Cyberattack Calculator

How likely is a $10B, $100B or $1T cyberattack, and what would a year of them cost? One line from today’s world, with about two billion-dollar attacks a year, to a swarm era with hundreds. Set the count, how fast large attacks thin out and how many are tied to a frontier model, then read the chance of each size, the yearly bill and the chance that one forces an AI slowdown.

Pack 030 asked how big a cyberattack would have to be to slow the AI frontier, and ran the numbers twice: for three worlds near today’s, with 1.8 to 9 billion-dollar attacks a year, and for three where such attacks are routine, with 100 to 250. This calculator puts both on one line. Start from any of the six worlds or set your own: how many billion-dollar attacks a year, how fast the large ones thin out, how many are tied to a frontier model, and over how many years.

The worlds are judgments, ours or yours, not forecasts. The record they are anchored on, and its sources, are in the pack. The attacks and outages reported lately are listed below the calculator.

Start from Attackers gain
Today’s rangeRoutine attacks
Three times as many billion-dollar events, with a slightly fatter tail: autonomous exploitation of known flaws outruns patching, as in October.

Chance of at least one, within 2 years

A $10B attack
74%
one every 18 months
A $100B attack
16%
one every 12 years
A $1T attack
2%
one every 93 years
A forced AI slowdown
13%
if a $100B attack tied to a model forces one 70% of the time

In a typical year of this world the largest attack costs about $9.8B, and the billion-dollar attacks together cost about $22B, 0.1% of US GDP.

Chance of at least one attack this size or larger, within 2 years

By size, on a logarithmic scale from $1B to $10T. Pink is the world you set. Grey is the 2023–26 record carried forward and black is the swarm era, the two ends this calculator runs between. The dashed line marks $100B, the size pack 030 takes as forcing a slowdown.

Defenders keep paceSwarm eraThis world$100B
0%25%50%75%100%$1B$10B$100B$1T$10TThis world: 74% chance of a $10B event within 2 yearsThis world: 16% chance of a $100B event within 2 yearsThis world: 2% chance of a $1T event within 2 years74%16%2%Swarm eraThis worldDefenders keep pace
0%25%50%75%100%$1B$10B$100B$1T$10TThis world: 74% chance of a $10B event within 2 yearsThis world: 16% chance of a $100B event within 2 yearsThis world: 2% chance of a $1T event within 2 years74%16%2%Swarm eraThis worldKeep pace

What a year of it costs

Attacks of $1B or more only. The yearly totals come from 2,000 simulated years, with no single attack counted above $10T.

Largest attack in a typical year
$9.8B
about the size of NotPetya; half of years see a larger one
A typical year’s bill
$22B
0.1% of US GDP
A 1-in-20 year
$185B
0.6% of US GDP

The six worlds side by side

Chance of at least one attack of each size and of a forced slowdown, within 2 years, and a typical year’s bill. Each world uses its own share tied to a model. The row in pink is the one on screen.

World$1B attacks a yearTail$10B$100B$1TSlowdownA typical year
Defenders keep pace1.81.0030%4%<1%<1%$4.2B
Attackers gain5.40.9074%16%2%13%$22B
The worm era returns90.7596%43%10%34%$66B
Commodity attacks1001.20>99%55%5%78%$399B
Ultra common1000.90>99%96%33%97%$910B
Swarm era2500.75>99%>99%94%>99%$7T

What has happened lately

The first slider is a count, so here is the count so far. These are the 39 attacks and outages logged in the 60 days to November 20, newest first, from the cyberattack and outage tracker: 34 that clear its bars and 5 smaller ones at the world’s largest companies. The tracker keeps all 107 major incidents since January 2026, the counts by week and month and the rules for what gets a row. 14 were reported in the past 14 days, 3.3 times the weekly pace before them. One in the past twelve months has a published loss of $1B or more (Binance); the slider’s lowest preset is 1.8 a year.

WhenIncidentWhat happenedScaleSource
Nov 19European CommissionStoppedAI namedThe Commission confirmed an attack it detected on October 22, after Politico reported it. Its spokesman called it an AI-generated campaign that used a publicly available tool. Staff HR and medical-insurance systems were offline for more than ten days; no leak of staff data has been found and no attacker named.Internal systems offline for more than 10 daysInfosecurity Magazine
Kyiv Post
Nov 17Rakuten SecuritiesTakenThe Japanese online broker said someone reached a test system on November 11. Login IDs and customer codes for up to about 11.25 million accounts may have leaked, and names, addresses, bank details and asset records for up to 4.47 million. Trading passwords and PINs were not taken, the company said.Up to 11.25M accountsThe Japan Times
OODA Loop
Nov 17Apple servicesOutageLargest co.Apple · below the barsAn internal update knocked out iCloud sign-in for about four and a half hours and the App Store for more than three. Apple’s own stores had trouble taking payments, and Apple Pay users reported failures. Apple ruled out an attack.About 4.5 hours for iCloud sign-inTechRepublic
inkl
Nov 16OptimizelyMany victimsLargest co.HSBC, Comcast, KrogerAn attacker exploited a flaw in Optimizely’s web-testing product to put malicious code on the sites of 178 organizations, among them the BBC, CNN, Zoom, HSBC and Comcast’s Peacock. It tried to swap cryptocurrency addresses and skim payment cards. Optimizely confirmed the incident.178 organizationsReport URI
Malwarebytes
Newsweek
Nov 16exploit_agentMany victimsAI namedGoogle reported the first attack pipeline it has seen run in the wild with no human in the loop. It compares web-application code before and after each patch, drafts an exploit and runs it. Targets were mostly French media, entertainment and education sites.Fully autonomous, per GoogleAI Weekly, on SiliconANGLE
Nov 16JPMorgan ChaseOutageLargest co.JPMorgan Chase · below the barsChase customers had trouble with online banking for about seven hours, and payments to and from other banks, including over Zelle, were delayed. The bank blamed a software bug in its own hardware systems and said nothing pointed to an attack.About 7 hours; 20,000+ Downdetector reportsCNN
AppleScoop
Nov 16A supplier to Visa, Mastercard and JPMorganTakenLargest co.Visa, Mastercard, JPMorgan Chase · below the barsThe extortion group n0n listed the three firms on November 16 and published files three days later. Each confirmed that data held by an outside provider was exposed; none reported a breach of its own core systems. The files hold marketing data: about 16,000 JPMorgan email addresses by a researcher’s count, and records on more than 6,000 Visa premium cardholders. Blamed: n0n, an extortion group.Thousands of customer records, through a supplierOgun Security
Cybernews
Nov 15AWS, Northern VirginiaOutageLargest co.Amazon · below the barsSign-in and identity services in AWS’s Northern Virginia region failed for 3 hours 16 minutes, with 11 services affected. AWS has not published a cause.3 hours 16 minutes, 11 servicesAppleScoop
RCR Wireless
NovA Japanese email-distribution serviceTakenAttackers compromised an email-distribution service used by about 400 organizations, exposing about 77 million addresses. Osaka’s city government alone lost more than a million.About 77M addresses at about 400 customersCheck Point Research
Nov 14Maileva (La Poste)TakenLa Poste said an attack in mid-September on Maileva, its business-mail subsidiary, exposed the names and postal addresses of 22 million people who receive mail from Maileva’s corporate clients. The data-protection regulator was notified and a criminal investigation is open.22M peoplefranceinfo
CNews
Nov 13WorkdayOutageWorkday, the payroll and HR software used by large employers, was disrupted for about 19 and a half hours before most customers were restored. A disaster-recovery test had left the backup site connected to live databases.About 19 hours 36 minutesAppleScoop
RCR Wireless
Nov 13KARTE, a Japanese marketing platformTakenUnauthorized access to the customer-analytics platform run by Plaid exposed data it held for client companies: about 19.11 million records at DHC, up to 18.29 million at JINS and about 6.5 million at MUFG Bank.19.11M records at one client aloneGlobal Cosmetics News
BigGo Finance
Nov 13FAA air traffic control networkOutageInstability in the FAA’s network halted departures across the US for more than four hours. A router had flooded the network with connection attempts during a copper-to-fiber upgrade; the investigation continues.A nationwide ground stop of more than four hoursMaintech
Nov 13Anthropic’s ClaudeOutageLargest co.Anthropic · below the barsSign-in to Claude failed for two hours, and later the same day the app, the API and Claude Code were down for about two and a half. No cause was given. Separately, Claude’s Microsoft 365 connector was out for a day with Microsoft’s own outage.Two outages: 2 hours, then 2 hours 29 minutesAppleScoop
Nov 12GitHub Actions secret stealerMany victimsFrom November 10, an attacker used 4,837 hijacked GitHub accounts to push code that steals build secrets and cloud credentials, opening probe requests in 61,764 repositories. It ran in at least 1,835 of them, according to Aikido, which named the campaign wfx-loot.Secrets stolen from at least 1,835 repositoriesBleepingComputer
Help Net Security
Nov 12Microsoft 365OutageLargest co.MicrosoftOutlook, Teams, Copilot and Exchange Online failed around the world for about a day. The cause was a software bug introduced in July, exposed when Teams traffic failed over to servers already short of capacity; Microsoft found no malicious activity. Downdetector logged more than 7.4 million reports.About 24 hours; 7.4M+ Downdetector reportsBusiness Standard
Maintech
Nov 12Autonomous agent campaignTakenAI namedPalo Alto Networks’ Unit 42 described a campaign run by AI agents built on DeepSeek and Zhipu GLM models: 25 operations in August and September against 19 organizations, 10 of which were breached, exposing more than a million records.10 organizations breached, 1M+ recordsPalo Alto Networks Unit 42, via DarkOwl
Nov 10AntiFox agent intrusionsTakenAI namedArctic Wolf tied a set of intrusions to AntiFox, a fork of the ARTEX attack system that its operators host themselves. The agents planted webshells on Microsoft IIS servers, read production secrets from AWS within about half an hour and deleted stored data.AI agents, per Arctic WolfArctic Wolf, via AgentBeam
DarkOwl
Nov 9NitoriTakenThe Japanese furniture retailer said an attacker who exploited a flaw in an administrative server may have taken up to 37.62 million customer records. The figure is a ceiling and includes duplicates.Up to 37.62M recordsMLex
Security Boulevard
Nov 5NTT Docomo’s 5G networkOutageDocomo’s standalone 5G network failed across Japan three times in eleven days: for about 14 hours from November 5, about nine on November 11 and nearly six on November 16. Calls, data and texts were hit for customers on the service. Docomo found no sign of an attack.About 14 hours nationwide, then twice moreTelecompaper
ASCII.jp
Jiji Press, on Nippon.com
Nov 4Crédit Mutuel ArkéaTakenThe French bank said attackers used a service provider’s platform to send fraudulent transfers from the bank’s own account. Le Télégramme put the sum at about €118 million ($133 million). The bank accepted the order of magnitude without confirming a final figure, and said part of the money had been blocked.About €118M ($133M), per press reportsL’Usine Digitale
Le Monde Informatique
Oct 30ChainalysisTakenAI namedThe blockchain-analysis firm said outside attackers reached internal systems in early October using agentic AI techniques, exposing some customers’ names, emails and job titles.Customer contact detailsStrobes
Loginsoft
Oct 29BinanceTakenAI namedAbout $6.36 billion was moved out of Binance’s wallets on ten blockchains, most of it in about 20 minutes: the largest theft on record. The FBI blamed North Korea. About $3 billion was frozen. On the way in, the attackers abused one of Binance’s internal AI tools. Blamed: North Korea’s TraderTraitor, per the FBI.$6.36B stolenTRM Labs
SOCRadar
Oct 26University of New Mexico Health Sciences CenterStoppedRansomware led the hospital to restrict admissions, divert ambulances and suspend outpatient surgery.Ambulances diverted, surgery suspendedBleepingComputer, via Cyber Management Alliance
Oct 23South Korean organizationsTakenAI namedPolice referred to prosecutors a former penetration tester whose group breached more than 90 South Korean organizations and stole about 8.2 million records. From late August he paired a coding agent with the ARTEX AI system to breach about 60 of them.More than 90 organizations, about 8.2M recordsSOCRadar
Fortune
Oct 22CapcomTakenUnauthorized access to the game maker’s membership service exposed 11.77 million records, mainly email addresses and login history. Passwords on 196 accounts were changed.11.77M recordsSOCRadar
Oct 21PHP emergency (PHP2Shell)Many victimsAI namedMaintainers shipped emergency fixes for two critical flaws in PHP, the language behind most websites, after AI worked out how to exploit bugs long thought too hard to use. Attempts were seen from October 17. The NSA told organizations to update at once or go offline, and government websites went offline.An order to patch at once or disconnectTechCrunch
SOCRadar
Oct 19Google Cloud, LondonOutageLargest co.AlphabetWith one power feed down for maintenance, a fault cut the other, and a misconfigured controller kept the generators from starting.6 hours 8 minutesControl Plane
Oct 19Cairn AI-agent break-insMany victimsAI namedA Chinese-speaking operator used AI agents to try more than 1,200 websites and break into more than 100 companies, for about $5,000 of model usage. Blamed: A Chinese-speaking operator.More than 100 companiesSecurity Boulevard
Loginsoft
Oct 16ConnectWise Cloud BackupMany victimsAI namedAn attacker abused the Microsoft 365 credential behind the backup service to sign in to customers’ tenants: attempts in almost 1,000 and success in more than 300. Microsoft blamed the China-aligned group Storm-0940, which used the ARTEX AI penetration-testing framework. Blamed: Storm-0940, China state-aligned, per Microsoft.More than 300 Microsoft 365 tenantsSOCRadar
Fox News
Oct 16Burger King JapanTakenAbout 11.18 million customer records were accessed between October 2 and 7.About 11.2M customer recordsBleepingComputer, via Cyber Management Alliance
Oct 12French supermarket refrigerationStoppedOvernight on October 11 to 12, attackers forced remote refrigeration controllers at about 700 stores into defrost cycles. At least 15 fires were reported and food losses ran to tens of millions of euros. Carrefour, Intermarché and Coopérative U were among the chains. Blamed: An IRGC cyber unit, per the French government.About 700 storesSOCRadar
Business Standard
Oct 12Ahold DelhaizeStoppedThe supermarket group took systems offline after finding malicious activity. Websites, apps and deliveries were disrupted at Albert Heijn, Delhaize, Stop & Shop, Food Lion and Giant; stores stayed open.Online sales and deliveries at five chainsBusiness Standard
Loginsoft
Oct 8Google Cloud StorageOutageLargest co.AlphabetWrites to Google Cloud Storage were disrupted for a day and six hours, by the duration in Google’s incident report.1 day 6 hoursPresenc AI’s cloud outage tracker
Oct 7IDC Frontier’s cloud (SoftBank)StoppedRansomware got administrator rights over the Japanese cloud’s virtualization platform and encrypted customers’ virtual machines, disrupting 495 companies and local governments. No ransom was paid.495 companies and local governmentsSOCRadar
Sep 30Microsoft Azure, several regionsOutageLargest co.MicrosoftPaused gateway updates all resumed at once and overloaded the service that manages regional networking.5 hours 45 minutesControl Plane
Sep 29Azure, Sweden CentralOutageLargest co.MicrosoftAzure’s Sweden Central region, including the Azure OpenAI service there, was disrupted for nearly six hours.About 5 hours 55 minutesPresenc AI’s cloud outage tracker
Sep 24BitgetTaken$387.5 million was moved out of the exchange’s hot and warm wallets, after a zero-day in third-party security products gave attackers privileged internal logins. Blamed: North Korea suspected; no government attribution.$387.5M stolenSOCRadar
StealthEX
Sep 22FBI recruitment siteTakenShinyHunters claimed it broke into the FBI’s jobs site through an unpatched Oracle PeopleSoft flaw, took 2 to 3 TB of data on employees and applicants and knocked the site offline. The FBI said a contractor had failed to apply an available patch. Blamed: ShinyHunters claimed it.2 to 3 TB claimedSOCRadar
CSIS

All 121 rows since January · the count by week · the largest companies · the log as a file.

What the calculator says

One line runs from today to the swarm. At the 2023–26 record of 1.8 billion-dollar attacks a year, with the tail at 1.0 so that each tenfold step in size is 10 times rarer, a $100B attack comes once every 56 years and a $1T attack once every 556 years. At 250 a year with a tail of 0.75, the swarm era, they come about 8 times a year and about 1.4 times a year.

The count moves the small sizes and the tail moves the large ones. With 100 attacks a year and a thin tail of 1.2, a $1T attack comes once every 40 years. With 9 a year and a fat tail of 0.75, it comes once every 20 years. If the question is about the largest attacks, move the tail before the count.

With the tail below 1, the yearly bill grows faster than the count. Attackers gain and ultra common share a tail of 0.9. With 18.5 times the attacks, a typical year costs 41 times as much: about $22B of billion-dollar attacks in the first and $910B in the second, 2.8% of US GDP. In the swarm era it is $7T, 21% of US GDP. An economy cannot lose that much and stay wired the same way, so somewhere on that road the line has to break. The calculator does not know where.

A forced slowdown stops being a tail risk long before the swarm. Within 2 years the chance is under 1% if defenders keep pace, 13% if attackers gain and 34% if the worm era returns. In the three routine-attack worlds it runs from 78% to over 99%.

The third slider carries the most doubt. Only an attack tied in public to a frontier model can force a slowdown, and pack 030 found none among this year’s large losses. If the attacks of your world run on open models that no US rule reaches, set the share lower. If they run on a US lab’s model through a jailbreak, set it higher. At zero the chance of a slowdown is zero in every world.

What the line leaves out. It has no ramp: the rate applies from the first day, which brings the risk forward. And the bar would move. A $100B loss every few months is a different shock from the first one, so the 70% is too high in the worlds where it matters most.

How the calculator works

The line. Attacks of $X billion or more arrive at the rate of $1B attacks × X−tail a year, and the chance of at least one within T years is 1 − e−rate × T. The tail is how fast large attacks thin out: at 1.0 an attack ten times larger is ten times rarer, at 0.5 about three times rarer.

The six worlds. Three are pack 030’s worlds for the next few years: defenders keep pace 1.8 a year with a tail of 1; attackers gain 5.4 a year with a tail of 0.9; the worm era returns 9 a year with a tail of 0.75. Three are its routine-attack worlds: commodity attacks 100 a year with a tail of 1.2; ultra common 100 a year with a tail of 0.9; swarm era 250 a year with a tail of 0.75. All six are judgments and none is estimated from data. The record behind them, ten attacks of $1B or more since 2017 and the catalogue of the worm years, is in the pack with its sources.

From attack to slowdown. A true slowdown is a binding limit on training or releasing more capable models, across the US labs, for six months or more. The chance of one within T years is 1 − e−share × rate × T × k, where share is the part of large attacks tied to a frontier model and k = 15% × (10−tail − 100−tail) + 70% × 100−tail: an attack of $10B to $100B tied to a model forces a slowdown 15% of the time and one above $100B 70%. Those two figures are the pack’s judgments. Each preset starts from the pack’s share, 5% + 60% × (multiple − 1) ÷ multiple, where the multiple is the rate over the 2023–26 record.

A year’s bill. The largest attack of a typical year is the median of the year’s largest, (rate ÷ ln 2)1/tail in $B. Below 0.69 attacks a year, more than half of years have none. The yearly totals add up the $1B attacks of 2,000 simulated years, with a Poisson count and sizes drawn from the same tail, each capped at $10T, about 30% of US GDP. The cap moves the 1-in-20 year when the tail is below 1 and barely touches the typical one. Losses below $1B are not counted. GDP is $33T.

Against the pack. The pack counts from November 16, 2026 to the end of 2027, 2028 and 2030, which is 1.12, 2.12, 4.12 years. Two years here comes close to its end-of-2028 figures. Its headline figures, 15% for a $100B attack and 11% for a slowdown, weight its three worlds 45%, 40%, 15%. This page shows one world at a time.

Hypothetical results. The figures are what the line implies for the world you set. They change with each setting and they are not a forecast.