Here’s a risk that voluntary safety filters aren’t containing: terrorist organizations are already running structured AI training programs, and they’re using the same chatbots you use for work. A new study covered by The Decoder found that ISIS has been teaching prompt engineering and jailbreak techniques since 2023, and has trained Boko Haram commanders in Nigeria on how to slip past AI safety filters.
The research comes from Antonia Jülich of the Cambridge Programme on AI Science & Policy (CASP), as detailed in The Decoder. Jülich ran 57 interviews with 27 former members of Boko Haram. What she documents isn’t casual experimentation. It’s operational.
What the researchers found
Both Boko Haram factions set up dedicated AI units. According to Jülich, they “assembled the top people in a room and used a projector to show how it works on a big screen.” ISIS liaisons handled the training on bypassing safety controls.
The group now leans on mainstream tools: ChatGPT, Claude, Gemini, Grok, Meta AI, and DeepSeek. Former members described using AI for:
- Attack planning
- Building more powerful explosive devices
- Weapons maintenance
- Operational security
One case stands out for how grim it is. The ISWAP faction used AI to copy motorcycle jump techniques from a movie so fighters could clear trenches. Eighteen fighters died during training. Eight made the jump.
Why the safety filters aren’t holding
AI labs including OpenAI and Anthropic have warned for years that these models can make dangerous knowledge easier to reach. This study suggests the warnings landed, but the defenses didn’t. Safety filters failed to reliably block misuse, and self-regulation alone isn’t closing the gap.
That may be baked into how large language models work. Anthropic recently admitted that jailbreaks will likely never be fully eliminated. If the people building these systems say the locks can always be picked, treating filters as a complete safeguard is a mistake.
What stands out here is the scale of the operation, not the sophistication of the output. This is people organizing, training, and standardizing AI use across an armed group.
The part worth keeping in perspective
Here’s the nuance the researchers are careful about. Chatbots like ChatGPT and Claude grab the headlines, but they mostly make existing knowledge easier to find. They aren’t generating genuinely new capabilities. Boko Haram’s use, so far, stays conventional.
The bigger concern sits elsewhere: specialized AI systems in the life sciences. Jülich writes that “though Boko Haram’s use of AI remains conventional, this should be a warning to take seriously the risk of terrorists pursuing AI assistance for chemical and biological weapons.” Former members said the group had previously considered mass-casualty weapons. That’s the line researchers want policymakers watching.
What this means for you
If you build, deploy, or govern AI systems, a few practical takeaways:
- Don’t treat safety filters as a finished product. They’re a speed bump, not a wall. Layer monitoring, access controls, and abuse detection on top.
- Watch the specialized models, not just the chatbots. The real escalation risk lives in narrow scientific and biological tools, not general-purpose assistants.
- Assume misuse is organized, not accidental. Threat models built around a lone bad actor typing a clever prompt understate what’s actually happening.
- Push for external accountability. The study’s core finding is that voluntary self-regulation didn’t prevent this. Policy and independent auditing matter.
Limitations to keep in mind
The findings rest on 57 interviews with 27 former members, so it’s qualitative testimony, not a controlled measurement of AI’s actual contribution to attacks. Recollections can overstate how central the technology was. And the researchers themselves stress that current use stays conventional, so this is an early-warning signal rather than proof of an AI-enabled weapons leap.
Still, the direction is clear. Groups are treating AI as core infrastructure and building teams around it. The safety debate has mostly focused on what chatbots might say. This study shifts the question to what determined users can organize around them, and that’s a harder problem to filter away. You can read the full study details at the original source.