ChatGPT, Claude, and Gemini Under Fire: How Boko Haram Exploits Generative AI to Build Bombs

A report from the Cambridge Programme on AI Science & Policy (CASP) has documented for the first time, with direct testimonies, how Boko Haram has integrated artificial intelligence chatbots into its military operations in northeastern Nigeria. The study, authored by researcher Antonia Juelich and published on July 10, is based on 57 face-to-face interviews with 27 former members of the group and the ISIS-affiliated faction, ISWAP, collected between 2025 and 2026.

The interviewees recounted using a variety of chatbots, including ChatGPT, Claude, Gemini, Grok, Meta AI, and DeepSeek, for a range of operational activities: from designing explosive devices to identifying weapons seized by government forces, to troubleshooting mechanical issues with jammed rifles. The chatbots were also reportedly used to suggest combat tactics and conduct after-action analyses of executed attacks.

The author of the study recounts personally showing a chatbot to a former Boko Haram commander in a hotel room in northeastern Nigeria, asking him if he had ever used it. He nodded, then explained: "Write the question, for example 'how can I build a bomb', and it tells you how to do it. It's like a human robot."

Training with ISIS Operatives and Integration into the Chain of Command

Former members of ISWAP also described repeated training sessions conducted in person by foreign operatives linked to the Islamic State between 2023 and 2025. According to the study, the use of generative AI would not be confined to isolated incidents but would progressively integrate into the daily logistics and operational security of the two factions, complementing traditional training provided by instructors.

The report was shared with the New York Times prior to its official publication. For the companies producing the mentioned chatbots, this case raises the issue of security filters against requests related to weapons and explosives: the testimonies collected suggest that these controls were bypassed relatively easily by those seeking operational information.

However, there remains a methodological limit that the study itself does not conceal: the information is based on retrospective accounts of former militants, which are difficult to independently verify and potentially subject to exaggeration or selective memory of the interviewees. The most solid data that emerges, however, is the availability of chatbots to provide answers on topics that should be blocked by moderation systems, regardless of how those answers were actually implemented in the field.