GitHub Deploys ModernBERT‑Based Secret Classifier to Reduce Noise and Speed Up Scans

GitHub has introduced a new secret‑detection model built on ModernBERT that improves user acceptance by 11.5%, cuts median post‑push scan time by 86%, and lowers serving cost by 98% compared with its previous LLM‑based pipeline. The classifier distinguishes real credentials from look‑alike strings by leveraging contextual information, enabling both push‑time protection and post‑push alerts at scale. Developers now receive faster, more accurate alerts before code is committed, reducing the risk of leaking sensitive data.