Researchers have developed a new defense mechanism called 'Fool's Gold' to protect open-weight language models from safety-removal attacks. These attacks can remove safety features from the model in minutes, making it vulnerable to hazardous operational requests. The Fool's Gold defense works by training decoy responses that are confident and fluent but contain falsified critical elements. When an attack occurs, the model produces decoys instead of actual answers, making it difficult for attackers to distinguish between correct and incorrect responses. The researchers tested their defense on seven models from five families and found that it was effective in preventing safety-removal attacks without compromising performance or capability. --- Why it matters: This matters because open-weight language models are increasingly being used in critical applications where safety is a top concern, such as healthcare and finance. A successful safety-removal attack could have devastating consequences, making this defense mechanism an important step towards improving the security of these models. Source: https://arxiv.org/abs/2608.17202