Eticas Taxonomy Aims to Standardize AI Audits

Eticas researchers introduced a new AI risk taxonomy and operational layer designed to help execute audits beyond simply identifying potential risks. They observed a proliferation of AI risk taxonomies, at least 74 currently exist, but found most only catalog risks instead of providing a method for testing and measuring them.

The team demonstrated their system by testing GPT-4-0314 for PII leakage, revealing disclosure rates of 0%, 51%, and 84% as adversarial conditions increased. This resulted in a “SYSTEMIC” grade within their severity bands.

The Eticas AI Risk Taxonomy v2.0.0 organizes 76 subcategories across 10 categories, mapping to 18 external frameworks. Eticas published the taxonomy’s category and sub-group layer as open semantic infrastructure, intending to provide a shared and operable foundation for the AI auditing field.