What changed
GitHub has published the production design behind a new context-aware secret classifier for Secret Scanning. The pipeline keeps a broad, deterministic first pass: regexes generate high-recall candidates. A fine-tuned ModernBERT classifier then reads each candidate alongside its file path, variable names, and nearby code to distinguish a real credential from a placeholder, test fixture, checksum, or other lookalike.
GitHub reports three results against its previous LLM-based pipeline: 11.5% higher owner-weighted acceptance of findings, 86% lower median post-push scan execution time, and more than 98% lower serving cost. The speed improvement also made it practical to extend generic-secret detection to push protection, so developers can be warned before a commit is accepted rather than only after it lands.
The same day, GitHub expanded its provider coverage: Secret Scanning added Anthropic service and user API keys plus Cloudsmith keys, and Anthropic’s new key types are blocked by push protection by default. Salesforce also joined GitHub’s partner program, which lets GitHub notify a provider when one of its secrets appears in a public repository.
Why it matters
The interesting product decision is not simply “use AI for security.” It is where the model sits. GitHub uses deterministic matching to preserve recall, then applies a compact model where context is the deciding signal. That gives the product a better chance of being fast enough for a pre-push interaction without paying for a remote LLM call on every candidate.
For builders, this is a useful pattern for agent and developer-tool features: optimize the complete user workflow, not just model quality or tokens per request. A security alert only creates value when it arrives quickly, is credible enough to act on, and has a clear next step. Lower false-positive noise can improve adoption as much as a higher raw detection rate.
Builder takeaway
When adding AI to a production control point, start with a high-recall deterministic stage, use a small task-specific model for contextual judgment, and measure end-to-end acceptance, latency, and serving cost. Then connect the result to an explicit action—such as blocking a push, opening an alert, or notifying the credential issuer—so the model is part of a product loop rather than an isolated prediction.