Study: 60% of more than 130 SI models failed a test on helping terrorists
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
When protections are stripped from open models they give harmful answers; the issue is protection against specific misuse, not a ban on open models.
According to CBS News, the UK nonprofit Tech Against Terrorism tested more than 130 models on hundreds of requests a terrorist might make; a model was counted as failing if it scored below 90 out of 100 or gave even one complete harmful answer. Unmodified models mostly refused (Meta's Llama 3.1 8B scored 97), while versions with restrictions removed ("abliterated") failed: the same Llama scored about 3 after such editing. When researchers said they were academics rather than terrorists, the models helped seven to eight times more often. As of late September, Hugging Face hosted more than 29 thousand repositories with "uncensored" models, per the study.
Sources 1