GitHub Ships a Purpose-Built Model for Detecting Leaked Secrets
GitHub deployed a purpose-built, fine-tuned model for detecting leaked credentials in code, moving beyond general-purpose language models for secret scanning.
Details
- What’s new: the model is context-aware, reading surrounding code to identify likely credentials “without generating code or prose,” including unstructured passwords that don’t follow recognizable token formats
- Why it’s different: previous approaches relied more heavily on pattern/token matching or general-purpose models; this one is specifically fine-tuned for the secret-detection task
- Rollout channels: existing customers with AI-detected password alerts are automatically upgraded to the new model; push protection checks are in private preview for organizations with GitHub Secret Protection (GHSP) or GitHub Advanced Security (GHAS); Copilot security review integration is coming soon in private preview; GitHub Enterprise Server support is planned for GHES 3.23 in public preview
- Billing: AI-detected password alerts remain free with GHSP/GHAS; the new push protection and security review checks will consume GitHub AI Credits
- No published benchmarks: GitHub did not disclose accuracy figures or false positive/negative rate comparisons against previous detection methods
What happened next
This is the third security- and control-focused Copilot release this week alongside local sandboxing’s GA and local-model discovery in the CLI, reflecting a broader push to harden GitHub’s AI tooling against credential leakage as agentic coding workflows generate and handle more code autonomously.