Boko Haram’s two-year leap into frontier AI: What a Cambridge study really tells us about terror and AI

It was never really in doubt that insurgent groups like Boko Haram and the Islamic State West Africa Province (ISWAP) had joined the rest of the country in figuring out what frontier AI models like ChatGPT and Claude could do for them. Bandits livestream their raids on TikTok now. Kidnap gangs negotiate ransoms over WhatsApp. It would have been strange, at this point, if insurgent groups sitting on captured smartphones and stolen SIM cards had somehow stayed offline while nearly everyone in Nigeria went digital. 

Yet, what nobody could answer with any confidence was to what extent they were actually using AI and to what effect.

For instance, in 2017, a Boko Haram commander wrote to his Islamic State handlers with an oddly mundane complaint. His men had captured a stash of Dragunov rifles from the Nigerian military and, despite combing through handbooks and instructional videos, simply couldn’t get them to fire. They gave up in what he called despair.

Nine years later, that problem no longer exists. According to a new working paper from the University of Cambridge’s Programme on AI Science & Policy, when Boko Haram fighters seize unfamiliar weapons today, they take them to a room, type the model number into a chatbot, and get back instructions on loading, servicing and firing them. That single contrast, buried deep in a 93-page academic report, says more about the trajectory of this problem than any headline about “terrorists using ChatGPT” ever could.

The study, authored by researcher Antonia Juelich, is built on actual fieldwork. Juelich and a research assistant conducted 57 interviews with 27 former members of ISWAP and Jamā’at Ahl as-Sunnah lid-Da’wah wa’l-Jihād (JAS) in Borno and Adamawa states across 2025 and 2026. Fifteen of the 27 had direct knowledge of their faction’s AI use; the rest, mostly lower-ranking members, hadn’t been given access at all, which is itself a telling detail about how tightly this technology is guarded internally and a hint that this was never meant to be a rank-and-file tool.

What the interviews describe isn’t casual experimentation. Both factions reportedly set up dedicated AI units, staffed by bomb-makers, weapons specialists and engineers pulled off combat duty specifically to manage chatbot accounts, run internal training sessions and field questions from commanders in the field. 

Access was restricted by rank, largely capped at the qaid level, with foreign operatives, described by one former commander only as “the white guys” from Libya, France and Arab countries, credited with delivering the original training and continuing to advise on prompting and on evading safety restrictions. 

How Boko Haram uses AI 

Fighters told Juelich they used chatbots to refine explosive designs, calculate payload weight for weaponised drones and, in one striking account, work out how to modify motorcycles to jump over military trenches, a manoeuvre apparently borrowed from a Hollywood stunt. 

Some framed dangerous queries as being for “a film” to get past safety filters and ran identical prompts across several platforms until one gave a usable answer. OpenAI, Google and Anthropic have each said such use breaches their policies and that safeguards are continually reinforced, but the testimony suggests those filters were, for a determined and coached user, an inconvenience rather than a wall.

This is where the “to what extent” question stops being academic and starts mattering to anyone tracking the security situation in the North-East. The timeline in the Cambridge report is not isolated; it overlaps almost exactly with a period in which Boko Haram and ISWAP have looked considerably more capable than they did five years ago.

Nigerian troops abandoned outposts as insurgents overran military bases on at least fifteen occasions in a single recent year, according to an Associated Press count, seizing weapons and killing soldiers in raids from Gajibo to Wulgo to the Nigeria-Cameroon joint base. 

Analysts describe a group that has grown more decentralised, more coordinated and increasingly reliant on modified commercial drones to drop explosives on targets that used to be out of reach. Borno’s governor has warned that military formations are being dislodged “almost daily without confrontation”, language that would have sounded alarmist a few years ago and now reads as a fairly accurate description of events.

It would be a hasty conclusion to draw a straight line from chatbot access to captured military bases, and the researchers behind these findings don’t do that either. Security analysts studying the resurgence point mainly to other drivers: a decentralised command structure that lets factions strike several locations at once; an influx of battle-hardened foreign fighters; a lull in the Boko Haram-ISWAP rivalry that freed up manpower; and continued logistical support from the Islamic State core. 

None of those explanations mentions AI. But none of them excludes it either, and that is the uncomfortable overlap; since AI chatbots became widely accessible, the group hasn’t just adopted them quietly in the background. It has grown bolder, better resourced and territorially more ambitious, at a pace that has alarmed even retired generals. 

Whether AI is a genuine accelerant layered on top of those older, better-documented drivers, or simply a convenient tool a resurgent group picked up along the way, is precisely the question this report opens rather than closes.

What happens as AI models get better?

That question isn’t going to get easier to answer, because the tools involved are not standing still. Every few months brings a materially more capable model, and the gap between what a chatbot could explain in 2023 and what it can explain today has widened considerably. If foreign trainers were coaching mid-ranking commanders on jailbreaking techniques using yesterday’s models, there’s no reason to assume today’s more capable systems, and tomorrow’s, will be any harder to manipulate without a fundamental rethink of how safety testing works. 

Juelich’s own recommendation is pointed: AI developers and safety researchers, she argues, need to work directly with people who understand how armed groups actually behave, because red-teaming a model in a lab tells you very little about what a trained, motivated insurgent unit will eventually get it to do in the field.

Also read: AI safety: Australia is stress-testing AI for blackmail and deception, while Africa is still playing catch-up

For the Nigerian armed forces and every other agency saddled with counter-insurgency planning, the practical implication is that this can no longer sit outside the security conversation as a curiosity. If insurgents can now set up an internal AI unit with the same seriousness they give their bomb disposal or logistics units, then disrupting that capability should become part of the same intelligence picture as tracking arms supply routes or foreign fighter movements, not a footnote about technology but a live front in a war that is, by most available evidence, already very deadly and has gone on for too long.