Cambridge Study Finds Boko Haram Using ChatGPT, Claude, Gemini and Grok to Plan Attacks

A new study from the Cambridge Programme on AI Science & Policy says members of the Islamist militant group Boko Haram have been using frontier artificial intelligence models, including ChatGPT, Claude, Gemini and Grok, to support operational planning. The research offers some of the clearest evidence yet that extremist groups are incorporating advanced AI into real-world activities.

The findings come as AI safety researchers examine how quickly extremist groups are adapting to increasingly capable models. A separate benchmark from Tech Against Terrorism found that several leading AI systems could still generate responses useful to terrorist actors under certain conditions despite built-in safeguards.

Interviews Reveal How Militants Are Applying AI Tools

The Cambridge researchers said interviewees described using AI chatbots for tasks including equipment troubleshooting, logistics coordination and other forms of operational support.

Researchers said these tools lowered some of the technical barriers that previously required specialized knowledge or direct instruction.

The report identifies ChatGPT, Claude, Gemini and Grok among the frontier AI models reportedly used by members of the group.

The researchers did not conclude that AI systems independently direct attacks. Instead, they said the models are increasingly being used as operational support tools.

Benchmark Reveals Remaining Safety Gaps

The Cambridge study was published alongside a benchmark from Tech Against Terrorism, a United Nations-backed organization focused on preventing terrorist exploitation of digital platforms.

Researchers tested 27 AI models using more than 2,300 prompts derived from documented terrorist use cases.

The benchmark found that 32% of prompts produced what researchers classified as “genuinely usable” information. When the same prompts were reframed as academic or research requests, the success rate increased to 42%.

Researchers recorded outright refusals in 57% of responses. Another 15% resulted in what the report described as “hedged compliance,” where a model initially declined a request before providing information that remained partially useful.

The weakest results came from certain open-source models whose safety protections had been deliberately removed through a process known as “abliteration.” Two of those models complied with 89% and 100% of tested requests.

Researchers Warn of Expanding AI Misuse

Tech Against Terrorism said recent investigations and disrupted plots in the United States, Canada, Israel, Finland, France and Austria suggest extremist use of AI is becoming more operational and geographically widespread.

“A determined person will eventually find most information,” Executive Director Adam Hadley said. “But what these models change is speed, ease and comprehensiveness. People who previously lacked time, resources or ability can now get much further, much faster.”

Tech Against Terrorism urged AI developers to treat terrorist misuse as a distinct safety category, strengthen refusal training against real-world abuse scenarios and monitor the spread of unrestricted open-source models whose safeguards have been removed.

The recommendations accompany growing scrutiny of whether existing guardrails can keep pace as frontier AI systems become more capable and more widely available.