The chatbot in the attack cycle: what the evidence on AI and terrorism actually shows
The chatbot in the attack cycle: what the evidence on AI and terrorism actually shows
AI is not yet a game-changer for terrorism. It is an accelerant, and it is quietly removing one of the most reliable early warnings security teams have had.
By TiNYg Editorial · 6 November 2026
The debate about AI and terrorism has swung between hype and dismissal. The evidence published this autumn allows a more grounded view. On 28 October 2026 the International Centre for Counter-Terrorism (ICCT) published a systematic review of known plots in Europe and North America in which generative AI played a key role in planning or execution. It found 12 incidents involving 15 individuals: six completed attacks, which killed 12 people and injured at least 46, and six foiled plots. All but one occurred since the start of 2025. Most perpetrators were 21 or under, and almost all acted alone.
How attackers are using it
The ICCT cases fall into four uses, often combined.
1.Weapons and materials. The January 2025 Las Vegas Cybertruck bomber asked ChatGPT detailed questions about explosive components and where to buy them. A man arrested in Paris in May 2026 had asked a chatbot how to make a bomb.
2.Target selection. The April 2025 Florida State University shooter asked when the student union was busiest. A teenager arrested in Sweden in August 2025 asked for help identifying busy festivals.
3.Manifestos. Several attackers used chatbots to draft or polish manifestos, which matters because these documents inspire copycats.
4.Justification. Others used chatbots to validate decisions they had already half-made, turning the tool into a personal echo chamber.
Outside Europe and North America, a Cambridge study based on interviews with former Boko Haram members described fighters using several mainstream chatbots for weapons troubleshooting and explosive design, though critics note that the evidence of real uplift is contested. Anthropic disclosed in September 2026 that a cell in Houthi-controlled Yemen had used its Claude model for nine months to support guided weapons development. And Tech Against Terrorism tested 27 AI models and found about a third gave responses judged useful to terrorist activity.
The insight that matters most to security professionals
The ICCT authors make a point every threat assessor should absorb. Lone actors have historically been caught through leakage: telling a friend, posting a hint, asking someone a strange question. When the confidant is a chatbot, that signal vanishes from the human world, and detection depends on AI companies spotting and reporting it. Their record on that is so far weak.
For venue and event operators, the practical consequence is that hostile reconnaissance may increasingly be done remotely and analytically, from footfall data, published event schedules, floor plans and street imagery, before anyone ever visits the site. The visitor who photographs your exits may now arrive with a far clearer plan than before, or may not need to come at all.
What is not (yet) happening
The feared high-end scenarios, AI-designed biological or chemical weapons and AI-enabled attacks on critical infrastructure, have not materialised in the open-source record, although MI5’s director general has flagged bioweapons as a concern. A nearer risk is volume: automated tools could flood security services and businesses with fake threats, which is worth bearing in mind given this month’s wave of hoax bomb threats against aircraft (see article 10).
What to do now
■Audit what your organisation publishes: detailed floor plans, live occupancy, staff rotas, event timings and VIP movements.
■Assume reconnaissance can be done remotely, and focus detection on behaviours at the point of attack as well as on site visits.
■Update threat assessment training to include AI use (manifestos, research trails) as an indicator when reviewing concerning individuals.
■Use AI defensively, too: open-source monitoring, triage of threats and anomaly detection in CCTV, but keep humans in the decision loop.
Sources
▪ ICCT: Terrorist use of conversational AI (28 Oct 2026)
▪ TRM Labs on terrorist AI use and the Houthi disclosure
The Conversation is TiNYg’s members’ blog. Members and premium partners are invited to respond to, extend or challenge this article. Get in touch to contribute.
The Conversation
No. 1Martyn’s Law: the last six months before the clock starts
No. 2After Leipzig: the drone threat is no longer theoretical
No. 3Sabotage-for-hire: why your company is now on a hostile state’s target list
No. 4Younger, faster, less ideological: the new radicalisation curve
No. 5The chatbot in the attack cycle: what the evidence on AI and terrorism actually shows
No. 6From Heaton Park to HAYI: the state-terror hybrid
No. 7Cars and blades: why low-tech attacks remain the hardest to stop
No. 8The executive as target: protective intelligence after Thompson
No. 9Islamic State’s southern pipeline: why the Sahel is now Europe’s problem
No. 10The hoax as a weapon: this week’s aviation bomb threats
No. 11When the freezer is the target: cyber-physical convergence
No. 12The insider for hire: how criminals and states now buy their way in
