Amazon plans changes to internal AI safety tests after independent review

AVERI says its review of Amazon’s internal AI safety evaluations found topical diversity and consistency with Amazon’s public statements, but raised concerns about straightforward attack prompts, refusals that may rely on superficial cues, predominantly English prompts across risk categories and references that could become outdated. According to AVERI, Amazon reports it will refresh test sets, introduce tiered difficulty calibration, expand multilingual coverage and review prompts quarterly.