Untrusted text can contain instructions aimed at overriding a prompt or exposing hidden context. This scanner highlights several common wording patterns and shows where they occur. It is a review aid for pasted documents, retrieved web content, and user messages—not a security boundary or a replacement for isolation and permission checks.
Scan untrusted prompt content for common instruction-override, data-exfiltration, and role-impersonation patterns with line references.
Benign security discussions can trigger matches, while novel or obfuscated attacks can pass unnoticed. Treat findings as signals, not verdicts.
Load the sample, run the tool, and compare the report with the visible markers or values in the input. The output is deterministic for the selected options, so the same text produces the same report.
Each line is tested against a small set of transparent, case-insensitive rules covering override language, hidden-prompt requests, secret extraction, role impersonation, and external transmission. Text remains in the current browser page during processing.
Benign security discussions can trigger matches, while novel or obfuscated attacks can pass unnoticed. Treat findings as signals, not verdicts.
No. The processing in this page runs in your browser. Uploaded files are read locally by the page.
Yes, but review the result first and test it with the exact model, template engine, or schema used by your application.
The tool shows an input error instead of producing a misleading report. Tool-specific parsing errors include a short correction hint.
No. It checks the visible text using the stated rules. Model behavior, source quality, and application security still require separate review.
Yes. The page accepts Unicode text, although pattern matching and token estimates may vary for scripts and formats outside the examples described.
| Item | Details |
|---|---|
| Processing | Local browser JavaScript |
| Input | Plain text or supported structured text |
| Exports | TXT, JSON, and tabular CSV when available |