GitHub adopts cheaper AI model to catch secrets developers leave in code

GitHub trained a credential classifier built on ModernBERT to improve generic secret detection and cut false alerts. Median post-push scan time dropped 86% and model serving costs fell over 98% versus its previous LLM pipeline.

• Precision rose 13% and recall 19% in offline evaluation
• A/B test showed 11.5% higher alert acceptance
• Model exported to ONNX with 8-bit quantization
• Generic secret detection now extends to push protection