Methodology
Tech Meltdown Tracker gives every incident a magnitude from 1 to 10. The scale is logarithmic, like the one used for earthquakes. Each whole number is ten times the damage of the one below it.
No claim without a citation
The site states nothing it cannot point to a source for. Three rules enforce that.
- Every paragraph cites a source. Each paragraph on an incident page ends with numbered markers that link to the sources it rests on, listed at the foot of the page. A judged magnitude cites the sources behind its rationale the same way. The file format requires it: an incident with an uncited paragraph, or a citation to a page outside its own source list, does not load and the site does not build.
- Every figure carries a verbatim quote. Each number that feeds a magnitude is published beside the passage it came from, copied word for word, with a link to the page. A script fetches each cited page and fails the incident when the quoted passage is not on it. Pull quotes pass through the same check.
- A second AI agent fact-checks each incident before it is published. An AI agent researches and writes each incident. A separate AI agent that did not write it then reads every cited page, tests each sentence against it, and corrects or removes whatever the sources do not support. Incidents that allege fraud or misconduct also wait for the editor to review them before they appear.
One source is accepted when it is the company’s own statement, a regulator, a court filing or a major outlet. Otherwise we look for two that are independent. Claims that are contested are attributed. A number that only an attacker has claimed is treated as a claim until the company or a regulator confirms it.
Charts and totals are arithmetic on the incident files, so their source is the log itself. Each links to the log and the data file.
How magnitude works
Every reported harm converts to points. One point is one person’s data exposed. Other harms convert through the weights in the table below. An incident’s points are added together, and its magnitude is the base-10 logarithm of the total, rounded to one decimal and held between 1 and 10.
Because of the logarithm, a magnitude 8 is not twice a magnitude 4. It is 10,000 times the damage. Two incidents of magnitude 6 together make a 6.3.
| Kind of harm | Points | Magnitude 4 | Magnitude 6 | Magnitude 8 |
|---|---|---|---|---|
| People whose data was exposed | 1 per person | 10,000 | 1M | 100M |
| Users knocked offlinefor 24 hours | 0.02 per user, per hour offline | 20,833 | 2.1M | 208M |
| Money stolen or lost | 0.1 per dollar | $100,000 | $10M | $1B |
| Deaths | 1,000,000 per death | under 1 | 1 | 100 |
| Injuries | 50,000 per injury | under 1 | 20 | 2,000 |
Each cell is the amount of that one harm that alone reaches the magnitude. An outage scores users multiplied by hours, so outage figures assume 24 hours offline. A single death scores 6.0.
Worked example
Casely recalls about 429,200 power banks again after a death is reported following its 2025 recall
- 1 death1 million points
- at least 6 injuries300,000 points
1.3 million points in all. The base-10 logarithm of that is 6.1.
The incident’s page quotes the passage behind each figure.
When nobody has counted
Some incidents do real harm that no source puts a number on. The research agent gives each of those a judged level from the scale below, writes down what it rests on and cites the sources for it. The fact-check agent reviews the call. Judged levels stop at 6, so an incident nobody has counted cannot outrank a counted catastrophe. 32 of 52 incidents on the log are judged.
Magnitude 7.0 Measured: a solid badge or bar.Judged magnitude 4.0 Judged: an outlined one.
Every badge and chart on the site follows that rule. A judged level is a reading of the reporting, not a quantity, so company totals, damage comparisons and the totals on the front page leave it out.
- 2An embarrassment confined to one product.
- 3Widely mocked, with no measurable harm.
- 4Real harm to customers that nobody has counted.
- 5Serious harm, or a real risk to people's safety.
- 6The ceiling for an incident with no reported figure.
A half level, such as 4.5, sits between two of these descriptions.
Where the incidents fall on the scale
The 20 measured incidents have a median magnitude of 7.0. The 32 judged ones have a median of 4.0. The gap says more about reporting than about harm. A figure tends to be published when it is large, as with a theft or a breach notice. Outages and AI failures rarely come with a count, so they are judged, and judged levels are capped.
This log is not a census. It leans toward incidents big enough to be reported with a number, so small incidents are scarce here.
MeasuredJudged, no reported figureHow magnitude works
Incidents in each half point of magnitude. Judged ones sit on top of measured ones. Source: the log · data
What counts as an incident
An incident is a specific event, with a date, where technology broke in public. Government and critical systems count, such as an agency’s network going down or an air traffic system halting flights. Every incident is filed under one of these kinds.
- Outages. Services that went down.
- Hacks and leaks. Systems broken into and data that got out.
- AI gone wrong. Models and agents that misbehaved.
- Bricked and broken. Updates, launches and devices that broke things.
- Physical safety. Hardware and vehicles that hurt people.
- Heists. Money stolen through hacks and exploits.
Left out: fines, lawsuits, layoffs and other business news, rumors resting on one anonymous source, ordinary product criticism, vulnerabilities nobody exploited, and anything whose only sources are vendor marketing.
The date is the day the incident began or became public.
Ongoing or resolved
Every incident says whether the failure itself is still happening, such as a service still down, an attack or a drain still running, or devices still bricked. An investigation or a lawsuit that continues after the failure stopped does not make it ongoing. A resolved incident shows the day it ended when a source gives one.
A claim that something is still broken goes stale, so an ongoing status shows the day a source last confirmed it. The site shows an incident as ongoing only if that day falls in the last seven days, counting the day the site was built, and otherwise says no fix has been reported since then. Like a paragraph, the status cites its source.
The front page
The front page leads with the incidents dated in the last seven days, counting the day the site was built. Under the count is a tally of what happened, such as how many went down and how many were hacked, and how many are still ongoing. The site does not compare the week with the rest of the year. Earlier months were researched after the fact and are less complete, so the comparison would mostly measure our coverage. When the last seven days hold no incident, the front page leads with the largest incident of the last month.
Each line carries a one-word stamp set by the incident’s kind. Outages read “down” and heists “robbed.” A breach reads “hacked,” or “leaked” when a company’s own mistake exposed the data. Safety incidents read “hurt” only when a source reports an injury or a death.
“Most outages” counts each organization’s incidents filed as outages. “Most hacked” counts break-ins that took data, including through a vendor, and any incident an attack caused. An incident counts for every organization it names, and a tie goes to whoever was hit most recently.
The totals in the year’s intro add up the figures sources report, each at its stated number. A kind of harm that fewer than 3 incidents report is left out.
Organizations, never individuals
The subject of every incident is a company or another organization, such as a government agency, and what happened to it. No individual is named anywhere on the site, whether an executive, an employee, a researcher, an official or a victim. People appear by role, such as “the company’s chief executive,” and a quotation is attributed to an organization or a role.
A story about one person’s private conduct, personal statements or departure is not an incident and is not filed.
Attention versus damage
Where one of an incident’s cited sources has a Hacker News thread, we record the points and comments on its most-discussed story. The match is by exact address, which misses threads that link to a different article about the same event. 5 of 52 incidents have a matched thread.
With 12 or more, the front page fits a least-squares line of the logarithm of that total against magnitude, then ranks incidents by how far they sit from the line. With fewer, the front page shows no chart.
Limitations
- The weights are judgment calls. No arithmetic makes a death comparable to a data breach. The weights are meant to make a magnitude 8 about as grave in one kind of harm as in another. Reasonable people would choose differently.
- Magnitude measures reported harm. If a company never says how many people were affected, the incident is judged and capped at 6, however bad it was.
- Reported figures are taken as given. “At least” and “up to” are scored at the stated number. Early counts are often revised, and the log follows the revision when we see it.
- Judged and measured levels are different things. They share a scale so the log can be read in one view, but a judged 4 and a measured 4 are not the same amount of harm.
- Company totals double count. An incident involving two companies counts in full for each.
- The log is incomplete. It leans toward incidents covered in English and toward large companies. An incident missing from the log is not evidence that nothing happened.
- Hacker News is one room. Its attention reflects what its readers care about, which is software and the companies that make it.
The full dataset, with the magnitude of every incident, is available as JSON.