A real-world blind spot in frontier AI safety just came into focus.
A new University of Cambridge field study, based on interviews with former Boko Haram and ISWAP members in north-east Nigeria, documents how these groups set up dedicated internal units to systematically use commercial AI chatbots (ChatGPT, Claude, Gemini, Grok, DeepSeek, etc.). They reportedly applied them to attack planning, weapons troubleshooting, and explosive design by often jailbreaking safeguards through disguised prompts.
The report uncovered dedicated teams, external training from foreign fighters, tightly controlled access (lower ranks kept out), and organizational adoption starting around 2023. The study captures former members’ perceptions of effectiveness, with caveats on verification, but the pattern of structured use is what stands out.
Most AI safety discussions focus on state-level threats or broad societal risks like disinformation. They often miss this middle ground; resourceful non-state armed groups iterating on commercial tools with patience and small teams. Current guardrails and threat models weren’t primarily built for this.
Key questions the report raises:
- How do we strengthen evaluations by including conflict/terrorism researchers, not just technical red-teamers?
- What about behavioral detection for concentrated, iterative misuse versus broad restrictions that hit legitimate users (students, aid workers, journalists) in affected regions hardest?
- Is this an outlier, or an early sign of wider proliferation?
This deserves thoughtful attention from individuals following AI development, policy, and global security.
답글 남기기