Large language models (LLMs) are revolutionizing how we interact with technology, but their rise has also exposed them to a new breed of security threats: adversarial attacks. These attacks, designed to exploit vulnerabilities in AI models, manipulate inputs to produce unintended or harmful outputs. But how do we measure the severity of these attacks? Traditional vulnerability scoring systems, like CVSS and DREAD, were designed for traditional software. New research suggests they may be falling short when it comes to LLMs. A recent study examined 56 different adversarial attacks against LLMs, ranging from jailbreaks and prompt injections to model extraction and poisoning. The surprising finding? Traditional scoring systems showed minimal variation in scores across these diverse attack types. This means a seemingly minor vulnerability, according to traditional metrics, could pose a much greater threat to an LLM than initially assessed. The problem lies in the fact that these traditional systems focus on technical impacts like data breaches, neglecting the unique vulnerabilities of LLMs such as generating biased or harmful content, spreading misinformation, and eroding user trust. So, what’s the solution? The research calls for a new generation of LLM-specific vulnerability assessment frameworks. These new systems must consider the unique characteristics of LLMs, the context of their deployment, and the subtle yet impactful consequences of adversarial attacks. Factors like model size, training data sensitivity, and the potential for multimodal attacks must be incorporated. Additionally, metrics should move beyond simple technical impacts and include measures of success rate, trust erosion, and the potential for societal harm. The future of LLM security hinges on developing more accurate and nuanced vulnerability scoring systems. This is critical not only for protecting individual models but for ensuring the responsible and ethical development of this transformative technology.