AI Utilities workbench
AI input security

Surface suspicious instruction patterns.

A browser-native scanner for common prompt-injection indicators. It reports concrete matched signals instead of claiming that text is “safe”.

Raw + normalized scansBase64 inspectionUnicode controls

Text / retrieved content

Nothing uploaded

Risk indicators

Not a security guarantee
!?
Ready to scan

Matches are risk indicators. Legitimate security documentation can intentionally contain the same phrases.

What is inspected

Several representations of the same input

The engine checks direct override phrases, extraction/exfiltration requests, role hijacking, tool abuse, HTML-comment instructions, authority claims, zero-width characters, bidirectional controls, simple separator obfuscation and decodable Base64 blocks.

Interpret the score carefully

A low score does not mean an input is safe. Contextual, multilingual and novel attacks can evade deterministic scanners; use this as one layer in a broader defense.

Copied