AI went fromhelper to operator.
In 2024, AI ran none of the 15 attacks we can document: every one had a human at the keyboard and AI assisting. In 2026 so far, AI ran 27 of 45 on its own or nearly so. This is the public, sourced record of 105 attacks since January 2023, each graded by how strong the evidence is.
- Incidents
- 105
- Confirmed
- 73
- Strong
- 29
- Suspected
- 3
Which models did attackers use?
Gemini, Claude and Claude Code lead. 43 of 105 name no model at all.
Incidents naming each modelhatched: reported by the model's own maker
Named once each
How the model is known
By the model's maker
n=105 incidents, 86 model mentions: one incident can name several models, so the bars add up past 62. 42 incidents were first disclosed by the model's own maker about its own platform, so these bars count disclosures, not which model attackers prefer.
Open-weight models or closed ones?
Closed, by far: 58 name a closed model, 7 an open one.
n=105. Open means downloadable weights (Qwen, DeepSeek, GLM, Kimi) or open-source tools; closed means a vendor's hosted model. 41 of the 53 closed-only cases were reported by the model's own maker, who can see its own platform and cannot see a model running on an attacker's machine. Years under 10 incidents show counts only.
Did an agent harness run the attack, and was it named?
33 of 105 ran through an agent harness. 26 name it.
One square per incidentgreen: open-source harness
By kind
Named harnesses
n=105. 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. 7 harness cases give the framework no name.
Which countries are the attackers attributed to?
China, then North Korea and Russia. 51 of 105 name no country.
Attributed country
Who the actor is
How the country is known
n=105. Countries are the attribution the source gives (a government, the model's maker or the investigating firm); this site makes none of its own. Dashed circles are attributions resting only on language or infrastructure clues. One incident can name two countries.
What did the AI actually do?
Malware dev most often. Agentic intrusion: 7 in 2025, 23 in 2026.
n=105, counted by half-year; a blank cell means zero incidents, and the last column runs only to 31 Oct 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.
Who got hit, and how sure is anyone?
Government and technology most. 73 of 105 are confirmed.
Sector of the targets
n=105. Up to three sectors per incident, taken from the targets the source names.
Every incident, with its sources.