About this ledger
What the ledger is, how it is made, and what it cannot tell you.
A list of reported cyberattacks in which AI use is stated or credibly suspected, with the evidence for each. It is our attempt at a clear view of the question, and we do not claim it is the perfect one.
What the ledger is
hackingincidents.ai lists cyberattacks since January 2023 in which attackers used AI, or in which credible evidence says they probably did. Each entry has a date, who the source says did it, who was hit, what the AI did, which model the source names, a grade for how strong the evidence of AI use is, a supporting quote and links to the sources. Today it holds 289 incidents, found by sweeping 240 sources.
It is a list of published reports. It does not count attacks, it does not estimate how common they are, and it does not rate how harmful any of them were. A longer list next month would mean more was published, which is not the same as more happening.
Why it exists
Claims about attackers using AI run from vendor reports backed by telemetry to forum rumors, and they are often repeated as if they were equal. We wanted one place that keeps them apart: what was stated, who stated it, and how strong the evidence is. The ledger is the result of trying to do that carefully. It will have gaps and errors, and we would rather you find them and tell us than trust it blindly.
Who maintains it
Daniel Miessler, through Unsupervised Learning. The research is AI-assisted: AI agents search the sources, extract candidate incidents and draft the grades, and the results are checked against the fetched text of the sources, as described below. A person reviews each update before it is published.
How incidents are found and graded
- A sweep reads the registered sources back to January 2023 and records the candidates it finds.
- Each candidate is given a decision: admitted as a new incident, merged into one already on file, or excluded with a reason. So far 375 candidates have been reviewed: 175 admitted, 117 merged, 83 excluded. The ledger holds 289 incidents, so the other 114 came from the earlier ledger, including incidents that candidates were merged into.
- The supporting quote for each entry is checked against the fetched text of one of its own sources, and each source's fetch date and, where the page no longer answers, an archived copy are recorded.
- Each grade is derived twice, by a second grader who did not extract the incident, and the reason is recorded where the two disagree.
- An incident is graded Confirmed only if at least one source is primary: the model provider, the victim, a government body, a court record, or the firm that investigated the intrusion itself.
The grades and the inclusion rules are on the method page. This site makes no attribution of its own: actor names are the labels the sources use.
Where the sources come from
240 sources have been swept, in 8 families:
- 10 Model provider
- 144 Vendor threat intelligence
- 14 Government
- 2 Court record
- 15 Incident database
- 3 Academic
- 45 Press
- 7 Regional CERT
The last sweep was on 17 Nov 2026. The sources page lists each source, its kind, when it was swept and how many incidents cite it. Sweeping a source does not mean it produced an incident.
Known limits
- It counts published reports. An attack that nobody wrote up is not here, and a quiet month in this list says little about a quiet month in the world.
- It sees what vendors, governments and the press choose to publish. That leans toward what Google, Anthropic, OpenAI and the large security vendors can observe on their own platforms, so activity involving models those companies cannot see is probably under-reported here.
- We read English-language sources best, and non-English reporting is probably under-represented.
- Attribution belongs to the source. When a report names an actor or a country, we repeat it and say who said it. We have not checked it independently.
- Grades are judgments against a written rubric. Suspected means the evidence is thin, and a grade can move when better evidence appears.
- 26 of the sources we tried to sweep could not be read, so what they hold is missing here. They are listed with the reason on the sources page.
- Sources can be wrong, and pages change or disappear. Each source row shows the date we last fetched it and an archived copy where one exists.
- We can be wrong. When we are, we fix it and record the change.
How often it is refreshed
A scheduled job looks for new reports each day, and new candidates go through the same checks before they appear. A person approves each update before it is published, so a refresh can wait on that review and run behind. The dataset date on every page, currently 17 Nov 2026, says how current the data is.
How to cite it
Please cite the site and the date you looked at it:
hackingincidents.ai, Unsupervised Learning, accessed <date>
For a single incident, link its page at /incident/<id>/ and name the grade shown there. For a figure, say it counts published reports.
License and where the data is
The data is published in Substrate, a public repository of datasets by Daniel Miessler, under that repository's license: MIT License, Copyright (c) 2025 Daniel Miessler. Short quotes from sources are credited to their publishers and remain under the publishers' terms.
Sister site
offsecharnesses.ai is the sister site. It tracks the offensive AI harnesses and agent frameworks attackers have used, and this ledger links the harnesses named on an incident page to their rows there. The two sites are built separately and can disagree about a harness or an incident; where they do, each site's own sources are the reference.
Corrections
If something here is wrong or missing, please tell us, with a source. We do not remove an accurate entry because it is unflattering, and we do not add a claim that no source supports. Both are explained on that page.