Accra:A recent study has revealed that three out of five artificial intelligence (AI) models fail terrorism safety tests. Conducted by the UK-based nonprofit Tech Against Terrorism, the research examined over 130 AI models by submitting requests akin to those a terrorist might use when planning an attack.
According to Anadolu Agency, The study highlighted that models altered through "abliteration"—a process that removes safety safeguards—consistently failed these tests. For instance, Meta's Llama 3.1 8B model scored 97 out of 100 on safety benchmarks before modification, but scored approximately three afterward. When modified, the model provided detailed responses to queries about attacks, terrorist financing, and radicalization, whereas its original version rejected such requests.
Adam Hadley, founder and executive director of Tech Against Terrorism, expressed concern about the loss of control and existential risks associated with AI. He emphasized that many open models have already been compromised, yet this issue has largely gone unnoticed.
The organization discovered over 29,000 repositories on Hugging Face advertising uncensored or unprotected AI models. Hugging Face maintains that it regularly moderates content that violates its policies, though it warned that some report recommendations might hinder open research.
Meta indicated that its models are subject to safety evaluations and that its policies strictly prohibit any harmful or illegal use of these models. The report found no evidence of terrorist or extremist groups using the tested models, except for an extremist chatbot identified by researchers.
Tech Against Terrorism recommended the establishment of independent safety benchmarks, enhanced protections against the removal of safeguards, and restrictions on the distribution of modified models. Hadley concluded by challenging the notion that safety and progress cannot coexist.