A recent report emphasizes the difficulty in accurately measuring the security of Artificial Intelligence (AI) systems. The core argument is that existing benchmarks fail to capture the complexities of AI security, especially emergent systemic properties. ### The Problem with Benchmarks The report, linked [here](https://berryvilleiml.com/docs/no-security-meter-ai.pdf), questions the effectiveness of simply maximizing security and privacy benchmarks to ensure AI security. It asserts that these benchmarks don't accurately reflect true AI capabilities, particularly when considering emergent properties like security. ### Lessons from Software Security Drawing parallels with software security, the report notes the evolution from black box penetration testing to whitebox code analysis, architectural risk analysis, and process-driven standards like the Building Security In Maturity Model (**BSIMM**). Given AI's potentially greater impact than software, the report suggests that similar measurement approaches might be applicable. ### A Call for Vigilance The report advocates for practical progress in AI security through improved data management and risk management by implementing robust assurance processes. However, it cautions that a definitive "security meter" for AI remains elusive, necessitating heightened vigilance in AI security practices.