Three in five AI models fail safety tests, UK study finds

A study by the UK-based non-profit Tech Against Terrorism has found that three out of every five artificial intelligence (AI) models failed specialised safety tests designed to assess their safety, BBC News reported. The study evaluated more than 130 AI models, submitting to them more than 100 questions and prompts of the kind a person planning a terrorist attack might make. Researchers noted that models that had undergone 'abliteration' — the process of removing safety guardrails built into models — completely failed all the tests. One worrying example was Meta's 'Llama 3.1 8B' model. With its safety features intact, the model scored 97 out of 100 on a safety benchmark, but after the safety features were removed, that score fell to 3. Researchers said that after the safety features were removed, the model provided detailed information on how to carry out attacks, how to raise funds for terrorism and ways that could facilitate radicalisation, while the original model refused such requests. Tech Against Terrorism founder and Executive Director Adam Hadley said that although there are concerns about the possibility of losing control of AI and the dangers it could pose, in reality this is already happening. He noted that the safety of open-source models has gone off track, but that this had not yet come to people's attention. The study found that more than 29,000 repositories advertising uncensored AI models, or models without safety measures, were present on the Hugging Face platform. In response, Hugging Face said content that violates its rules is moderated in line with those rules, but warned that some of the study's proposals could obstruct open research. Meta said its models are subject to safety evaluations and prohibit use for dangerous or illegal activities. The researchers said that apart from one extremist chatbot run by an individual, no sufficient evidence had yet been seen that terrorist or extremist groups are using these models. Tech Against Terrorism called for establishing minimum benchmarks for AI safety, taking strong measures to stop the removal of safety features and controlling the spread of modified models. Adam Hadley further said it is not correct to believe that safety and progress cannot be pursued together. At a time when open-source AI models are becoming widespread, experts note that strengthening safety standards across the entire industry to protect against dangerous uses is essential.