hackingincidents.ai
Incidents289Sources240Dataset2026-11-17Last sweep2026-11-17
Overview / Analysis
HI // Analysis

What the reports show

Patterns across 289 incidents drawn from published reports of AI-assisted attacks since January 2023. Each pattern carries what it cannot tell you: these are reports, shaped by who looks and who publishes, and the attacks nobody wrote up are not here.

P//Patternsn=289

Six patterns, read with care

What the reports show, the numbers behind each pattern, and its limits.

P1

A growing share of reports describe AI running attack steps itself

79 of 177 in 2026 so far, 13 of 80 in 2025, 0 of 27 in 2024

These are reports that describe AI running steps of the attack itself, on its own or semi-autonomously with an operator directing, rather than helping a person who did the work.

Read with careThe ledger holds 177 incidents from 2026 and 27 from 2024, so the comparison rests on a small early base and on how much more is published and collected now. It cannot separate more AI-run attacks from more reporting of them.

P2

21% of incidents were first disclosed by the maker of the model involved

61 of 289 first disclosed by the model's own maker; 93 by an AI or cloud vendor of any kind

Model providers publish threat reports about abuse they detect on their own services, and some of those reports describe many separate operations.

Read with careA provider sees only its own models, so the model counts here follow who publishes. They are a weak guide to which models attackers prefer.

P3

Reports name closed models more often than open-weight ones

108 name a closed model, 23 an open-weight one, of 289; 12 name both

Closed means a vendor's hosted model; open-weight means downloadable weights such as Qwen, DeepSeek or Llama.

Read with careAn open-weight model can run on an attacker's own hardware, where no provider would see or report it, so these counts cannot show how often open-weight models are used.

P4

Country links cluster in a few countries

81 of 116 incidents that name a country name China, North Korea or Russia (one names two of them)

Countries are the ones the source names; 14 of the 116 rest on language or infrastructure clues alone, not an attribution. This site adds none of its own.

Read with careEach country link is the publisher's own, and publishers track the actors they already watch. 173 of 289 incidents name no country, so this describes 116 incidents, not the whole ledger.

P5

Most incidents rest on one independent publisher

201 of 289 have one independent publisher; 57 have two or more independent ones; 31 are known only through an incident database

An incident database that only restates a report we already cite is not counted as a second publisher.

Read with careA single independent publisher can be right, and a widely copied report can be wrong. Still, weigh an incident with one independent publisher more lightly, and check its grade and sources.

P6

51% are graded Confirmed

146 of 289 Confirmed, 110 Strong, 33 Suspected

Confirmed means the source states AI was used and shows evidence, such as provider telemetry, recovered prompts or malware that calls a model.

Read with careThat evidence is often telemetry outsiders cannot check, so Confirmed is not independent verification. Two graders working blind agreed on 207 of 289 grades. The grade says nothing about harm done or how much the AI contributed.

00//Autonomyn=289

Reports describe AI doing attack work itself in 92 of 289

How much of the attack did the AI run, by year? Click a segment to filter the charts.

25 ran on their own; 67 were semi-autonomous, with an operator directing. All 92 date from Jun 2025 on.

n=289
2023
n=5
2024
n=27
2025
n=80
2026
n=177
01//Modelsn=289

Gemini is named most often, in 31 of 289, then Claude 28, ChatGPT / GPT 27. 170 of 289 name no model at all.

Which models do the reports name?

119name a model

Incidents naming each modelhatched: reported by the model's own maker

Named once each

By the model's makergreen: open-weight maker

Who says AI was used: the model's maker, a researcher, or inferencen=289

70The model's maker says so
138A researcher names it
27The victim says so
38Inferred from artifacts
16Not stated
n=289 incidents, 160 model mentions

One incident can name several models, so the bars add up past 119. 61 incidents were first disclosed by the model's own maker about its own platform, so these bars count disclosures, not which model attackers prefer.

02//Open or closedn=289

Reports name a closed model in 108 of 289 and an open-weight one in 23.

Open-weight models or closed ones?

Model makers report the abuse they see on their own platforms. None of them can see an open-weight model running on an attacker's own machine, so these counts cannot show how often open-weight models are used.

n=289
2023
n=5
2024
n=27
2025
n=80
2026
n=177
n=289

Open means downloadable weights (Qwen, DeepSeek, GLM, Kimi) or open-source tools; closed means a vendor's hosted model. 59 of the 96 closed-only cases were reported by the model's own maker. Years under 10 incidents show counts only.

03//Harnessn=289

Reports describe an agent harness in 65 of 289; 56 of those name it.

Did an agent harness run the attack, and was it named?

One square per incidentgreen: open-source harness

Open-source harnessVendor coding agentCustom frameworkNo harnessHarness named

By kind

Named harnessescoloured by kind

n=289

A harness is the software that drives the model through the attack: an open-source offensive framework (PentAGI, HexStrike, Strix), a vendor's coding agent run as the operator (Claude Code), or a private framework the attackers built. 9 harness cases give the framework no name. One incident can name several harnesses, and a vendor agent such as Claude Code can run inside a custom framework, so the named bars count mentions and can exceed their kind.

04//Attributionn=289

Reports link 36 of 289 incidents to China, then North Korea 33 and Russia 13. 173 of 289 name no country, and of the rest, 14 rest on language or infrastructure clues alone.

Which countries do reports link the attackers to?

Attributed by the sourceLanguage or infrastructure clue onlyCircle area = incidents · 173 with no country are not drawn
China: 36North Korea: 33Russia: 13Iran: 13Hong Kong: 6Japan: 3Taiwan: 2Cambodia: 2Kazakhstan: 1Netherlands: 1South Korea: 1Ukraine: 1Greece: 1Indonesia: 1Pakistan: 1Serbia: 1Brazil: 1Algeria: 1China 36North Korea 33Russia 13Iran 13Hong Kong 6Japan 3Taiwan 2Cambodia 2Kazakhstan 1Netherlands 1South Korea 1Ukraine 1Greece 1Indonesia 1Pakistan 1Serbia 1Brazil 1Algeria 1

Country named by the source

Named once each

Who the actor is

How the country is known

Attributed by the source102
Language or infrastructure only14
No country given173
n=289

Countries are the ones the source names (a government, the model's maker or the investigating firm), whether as an attribution or from language or infrastructure clues; this site adds none of its own. On the map, dashed circles are country links resting only on language or infrastructure clues, not attributions. One incident can name two countries.

05//Techniquen=289

Malware dev most often, in 89 of 289. Agentic intrusion: 7 of 80 in 2025, 72 of 177 in 2026.

What do the reports say the AI did?

H1 ’23H2 ’23H1 ’24H2 ’24H1 ’25H2 ’25H1 ’26H2 ’26to dateAll
Prepare
59204089
Get in
51745214
Run the operation
795532612
Cash out
572316
n=289

Counted by half-year; a blank cell means zero incidents, and the last column runs only to 17 Nov 2026. One incident usually involves several techniques, so the columns add up past the incident count. Click a technique to filter by it, or a cell for that technique in that half-year.

06//Targets and graden=289

Government (55 of 289) and technology (54) are the sectors reports name most. 146 of 289 are graded Confirmed.

Which sectors do reports name, and how strong is the evidence?

Sector of the targets

Three incidents or fewer

Evidence grade

Who disclosed it first

Read the method for what each grade requires.

n=289

Up to three sectors per incident, taken from the targets the source names.

07//Incidents289 rows

The incident list

The incidents in the ledger, newest first, with a search and links to their sources.

See the incidents (289) →

Any chart filter you set here carries over to the list.