GitHub details ModernBERT secret-detection classifier with faster scans
GitHub published a technical explainer on a new secret-detection classifier that uses regex candidate generation plus a fine-tuned ModernBERT encoder to judge whether a string is a real credential or a lookalike. GitHub says the compact model cut median post-push scan execution time by 86% versus its previous LLM-based pipeline and lowered serving cost by 98%, making generic secret detection practical in push protection as well as post-push scanning. GitHub also reports 11.5% higher acceptance and 86% owner-weighted alert precision, saying the efficiency gain let it return results before a push is accepted.
GitHub says the classifier's efficiency made it practical to extend generic secret detection to push protection, so results can be returned before developers push changes rather than only after code is committed.
October 13, 2026 | Updated October 14, 2026
|
9 minutes
- Share:
11.5%
higher acceptance
Owner-weighted alert precision
#GitHub#Microsoft#ModernBERT#GitHub secret scanning#GitHub push protection#Microsoft CoreAI