Input
Paste redacted user prompts, retrieved text or tool output. Choose relevant policy categories. Up to 100,000 characters per scan.
A transparent inspection layer.
Analysis
Each of 23 regex patterns contributes its fixed weight at most once. Three optional structure checks add heuristic signals. Scores are summed and capped at 100.
Review
Below 30: no/low flags. From 30: review suggested. From 70: high signal. These inherited demonstration thresholds are uncalibrated and do not enforce policy in a real system.
Preserved engine and examples
The supplied override, exfiltration, tool-hijack and obfuscation checks remain. Five original example texts and category toggles are included. Structural rules look for invisible characters, long base64-like text and a high non-ASCII ratio. Ordinary quotations, technical documentation and non-English text can produce false positives; obfuscated and novel instructions can evade detection.
Privacy and exports
Input and the latest 50 log entries remain in memory in this page. Reloading clears the session. Findings export excludes payload by default, with an explicit checkbox to include it. Exported text may be sensitive; use test data. The cart separately stores only whether one license is selected. No network scanner, model API or server logging is connected.
Future integration
Licensing and purchase fulfillment are provided as a separate seller-operated service. The service must be deployed and connected before production sales. This release does not include agent middleware, server-side enforcement, third-party integrations or independently benchmarked detection. The scanner alone cannot establish safety, detect every attack or authorize a tool action.
Open scanner →