Anthropic publishes LLM biosecurity evaluation framework — GPT-4 shows mild threat creation uplift

AI Impact Summary

Anthropic researchers have published an evaluation framework for assessing LLM-aided biological threat creation risk, finding that GPT-4 provides only mild accuracy improvements in threat creation tasks. This work establishes a baseline methodology for measuring dual-use biosecurity risks across LLM providers and suggests current models pose limited but non-zero risk amplification. The framework is designed to inform responsible disclosure practices and guide future safety evaluations as model capabilities advance.