Exposure scores, first introduced in 2023 by Eloundou et al., quantify the proportion of occupational tasks that can be assisted by large language models, such as GPTs. These scores have become a crucial factor in shaping the future of work debate. However, as these scores are applied in various contexts, their limitations become apparent, highlighting the need for a more nuanced understanding of their implications. The scores primarily focus on the capabilities of language models, without fully accounting for the risks and security concerns associated with their adoption. The rapid development of large language models, particularly in areas like DeFi, is expanding their capabilities while introducing new security risks. This raises important questions about the long-term viability of these models in sensitive applications, so what matters to practitioners is the need to carefully consider the security implications of relying on large language models in critical systems1.