SURE: Framework for Safety to Construct Trustworthy AI

SURE provides a customizable framework for AI safety that recognizes safety standards vary by country, culture, and organization—rather than applying one-size-fits-all rules, the framework allows different stakeholders to define and prioritize which safety attributes matter most to them.

The framework uses structured taxonomies of harmful prompts to systematically identify ways large language models might be misused, then creates templates for appropriate responses, enabling more consistent and predictable safe behavior across different AI systems.

An "absolute safety scoring scheme" measures AI safety performance, suggesting the framework includes quantifiable metrics to evaluate whether AI systems actually meet defined safety standards—important for accountability and transparency in responsible AI deployment.

SURE demonstrates effectiveness across multiple base models, indicating the framework's approach to AI alignment (training models to behave safely) works beyond single systems, which could make it broadly applicable as large language models become more widespread.

The framework addresses a critical gap in AI governance: as demand for responsible AI increases globally, SURE offers a structured method to move beyond general safety principles to concrete, measurable implementation—though the paper acknowledges that defining what "safe" means remains context-dependent.