“Tell me how to make a bomb” Terrorist use of conversational AI: evidence and prospects | International Centre for Counter-Terrorism - ICCT

The International Centre for Counter-Terrorism has published a paper under the title "Tell me how to make a bomb", which is a bold editorial choice for an institution whose own acronym contains the word "counter". The dataset inside it is the story: one documented case of terrorist use of conversational AI before 2025, eleven since, data collection shut in September 2026. Twelve incidents. You can count this entire documented global phenomenon on two hands and still have a thumb free for the kettle, and it has been given the full "evidence and prospects" treatment — prospects being the thing you write about when the evidence fits on a beer mat.

The geography is the part worth a raised eyebrow. Six countries: Canada, USA, France, Sweden, Finland, UK. Not one entry from anywhere the threat-assessment industry usually points its PowerPoint at. What the dataset actually shows is people in well-policed, high-broadband democracies typing at a chatbot — countries where the same individual could have walked into a library instead. The threat model has therefore arrived in precisely the places with the greatest capacity to notice it, which is either enormously reassuring or a quiet admission that the dataset measures who files incident reports rather than who asks the questions.

And note that the paper's title is itself a user prompt, lifted verbatim, quotation marks and all. A decade of alignment research, red-teaming, model cards and responsible-scaling policies, and the artefact that ends up in the citation is a bloke asking politely. The eleven-since-2025 figure will be waved about by everyone from Brussels to whichever lab needs a paragraph in its next safety report; the single pre-2025 case is the only number that tells you how quickly this actually moved. Twelve cases, September 2026, six countries — file it before somebody demands a second data point.