PROMPT INJECTION CHECK
How it works
A clear path from a selected input to findings, coverage and a local report.
- Select input
- Analyze locally
- Review findings
- Export results
From input to report
Choose text or a supported file. The browser reads its supported layers, prepares text for analysis, applies rule-based checks, then presents findings with locations and a coverage summary. You decide whether to save a report.
Local document processing
Selected documents are parsed in your browser. The scanner does not send their contents to a server for analysis. Public site resources and the separate visit counter may still make ordinary site requests.
More than visible text
Some files contain comments, alternative text channels, metadata or formatting that is not obvious in their normal view. Supported parsers inspect relevant layers when available. The coverage summary identifies layers the scan could not inspect.
Preparing content for checks
Supported text and structured values are prepared for local rule checks. Some encoded or fragmented representations can be interpreted within limits. This does not mean every encoding or every document layer is understood.
Detection and language
Checks consider document structure and instruction-like content. Every scan uses the available language rule sets, regardless of the interface language. The interface setting changes labels, not detection coverage.
Findings and coverage are different
A finding points to unusual or suspicious material with a location, excerpt and reason. Coverage describes what was inspected, partly inspected or unavailable. Read both before deciding how to handle the source.
Batch workflow and reports
Select up to 12 supported files in one batch. They are scanned sequentially in the browser. Review each result separately; export the selected result as JSON, TXT or Markdown, or print that result to PDF. Export all saves separate JSON, TXT or Markdown reports in one ZIP for successful files.
How detection improves
A new case is reproduced and turned into a controlled test before a rule change is considered. Regression tests compare missed suspicious cases with unnecessary alerts. Changes are reviewed across earlier and fresh cases before release; more findings alone are not the goal.
Human review remains important
The scanner helps you inspect untrusted content. It cannot establish an author’s intent, certify a document as safe or predict how a particular AI system will respond. Compare a finding with the original and keep the receiving AI’s permissions limited.