This research introduces an open-source testing framework to evaluate how large language models handle potentially dual-use scientific information.

• Creates reproducible benchmarks measuring both harmful content refusal and legitimate scientific discourse allowance
• Tests four major LLMs with systematically varied prompts related to controlled substances
• Reveals distinct safety profiles among different models, highlighting the tension between protection and censorship
• Provides valuable insights for AI safety implementation without over-restricting scientific knowledge

For security professionals, this research offers critical metrics to evaluate AI systems' ability to appropriately handle sensitive content while maintaining scientific integrity.

Forbidden Science: Dual-Use AI Challenge Benchmark and Scientific Refusal Tests