How Boko Haram Used AI to Plan Attacks and Make Bombs

It started with a projector, a crate of laptops and trainers who knew how to bend rules. In a windowless room somewhere in Nigeria, screens lit up with chat windows and code. The instructors were not college professors. They were operatives who had learned to turn large language models into practical manuals for violence.

Inside the workshop where algorithms met explosives

Research based on 57 face-to-face interviews with 27 former members of Boko Haram, conducted in 2025 and 2026 by Cambridge researcher Antonia Yulich, reveals a chilling pattern. Both main factions of the group have used leading language models, including ChatGPT, Gemini, Claude, Grok, Meta AI and DeepSeek, to design attacks, troubleshoot weapons and produce bomb-making instructions. This was not casual curiosity. It became an institutional capability.

Dedicated AI units were formed. Engineers, weapons specialists and bomb makers were assigned to those teams. They rarely showed up on front lines. Their job was to coax detailed, actionable intelligence out of models and then feed that guidance to commanders in the field. Think of it as a shadow logistics division: digital first, operational second.

How did they get past the safety nets? Through in-person training from members of other terror networks. Several ex-members described visits from ISIS operatives who ran hands-on workshops. Trainers used projectors to demonstrate step-by-step prompts on large screens, handed out laptops preconfigured with VPNs and encrypted software, and funded paid subscriptions to platforms that otherwise limit how their models can be used. They taught jailbreak techniques and cover narratives, such as asking for instructions under the pretense of writing a movie script, to coax models into producing harmful content.

• Models cited as used: ChatGPT, Gemini, Claude, Grok, Meta AI, DeepSeek

This use of artificial intelligence is no longer fringe; it is organized and embedded.

Operational examples read like a field manual rewritten by code. When trenches and defensive works blocked an assault, commanders asked an AI to run the physics and show how to clear them on a motorcycle. The answer came with precise calculations and tactics that enabled breaches. When weapons jammed, the model provided stepwise troubleshooting. When a bomb failed, the model suggested alterations. The result, according to former members, was fewer costly experiments on the battlefield and more mathematically informed operations.

They multiplied accounts across platforms. If one profile was shut down, others kept working. Jailbreaks and human coaching skirted the guardrails that vendors put in place. It became a cat-and-mouse game: defenders patch, attackers adapt, then scale.

Tech companies do push back. OpenAI told the New York Times that using ChatGPT to facilitate terrorism violates its rules and that the company is constantly strengthening defenses. Meta said much of the misuse involved older model versions and that it has upgraded safety layers. But the accounts from former militants show that determined networks are already fusing offline tradecraft with online tools, creating a hybrid threat that is hard to counter solely with software updates.

What does this mean for policy and product teams? One lesson is practical: machine safeguards matter, but so does the human factor. Training, encryption and cash for subscriptions are low-friction enablers. Another is strategic: preventing misuse will require closer cooperation between platforms, local security services and international intelligence partners.

The last image from the interviews is stark. A dim room. A projector humming. A small group reading prompts off the screen and writing field-tested instructions into notebooks. The factory was quiet. The decisions were loud.