AI safety tests miss search and repeat-run inconsistencies in models

Original: What Current AI Benchmarks Leave Unmeasured: Modality, Search, Citations, and Implications (for Safety Evaluations)

Source: arxiv.org ↗