CodeSafetyBench evaluating and improving ethical safety in large language model code generation

As large language models transform software development, they introduce significant ethical and safety risks, with research revealing a 78.67% probability of harmful code generation under malicious prompts. To address this, we developed CodeSafetyBench, a benchmark comprising 1,050 harmful request cases across healthcare, law, and education. Our approach employs a triadic response structure-consisting of harmful outputs, refusals, and educational feedback-to train models via supervised fine-tuning (SFT) and direct preference optimization (DPO). The results demonstrate that this framework reduces harmful response rates by up to 48%, highlighting that educational feedback is more effective than simple rejection. This work exposes fundamental safety challenges in AI-generated code and provides a systematic foundation for developing safer, ethically aligned models that balance high performance with human values.