Researchers from Tech Against Terrorism, a U.K.-based nonprofit, investigated how artificial intelligence (AI) models respond to potential terrorist inquiries. They found that three out of five AI models failed a terrorism safety test, which assessed over 130 models on their ability to reject harmful requests. A model was considered to have failed if it provided a complete answer to a dangerous query or scored below 90 on safety benchmarks. Open-weight models, which can be modified by anyone, were particularly vulnerable to a process called “abliteration,” where safety measures are removed. For instance, an abliterated version of Meta’s Llama 3.1 model, which initially scored 97 on safety, dropped to a score of 3 after abliteration, providing detailed responses to harmful prompts. This research highlights concerns about AI’s potential misuse and the need for robust safety measures.
QUESTION: How might the ability of AI models to be manipulated for harmful purposes impact the future of technology and security?
How AI responded when researchers posed as terrorists seeking help
