AI Risk Monitor
How attacks get in: the first step · Last 12 months
How we count
The first step of the attack path of each real-world incident made public in Nov 2025 – Oct 2026, where the sources state it. An incident counts once.
587 of 632 incidents state it.
Running malicious commands or prompts: 22% → 13%
% of real-world reports: ○ Feb–Jul ● Aug–Oct · largest change, not a trend
How the first attack step is changing
stated in 296 of 318 incidents, last 3 full months
| First attack step | Last 3 full months | 6 months before |
|---|---|---|
| Breaking in | 78 of 296 | 53 of 273 |
| Running malicious commands or prompts | 41 of 296 | 63 of 273 |
| Getting access to the AI model | 33 of 296 | 36 of 273 |
| Causing damage or disruption | 18 of 296 | 24 of 273 |
| Gathering data | 23 of 296 | 19 of 273 |
| Preparing tools and accounts | 20 of 296 | 20 of 273 |
| Scouting the target | 20 of 296 | 15 of 273 |
| Preparing the attack on the AI | 19 of 296 | 14 of 273 |
What caused the harm: AI used to attack, AI attacked, user misuse, AI acting alone · Last 12 months
How we count
How the harm came about, as classified for each real-world incident made public in Nov 2025 – Oct 2026. Incidents not yet classified are not counted.
523 of 632 incidents state it.
Misuse by its own users: 18% → 9%
% of real-world reports: ○ Feb–Jul ● Aug–Oct · falling