PROMPT / DOCUMENT / TOOL OUTPUTA closer look.
A closer look.
Before the next action.
Ondrex inspects text for instruction overrides, disclosure requests, tool-hijack phrasing and obfuscation signals. Transparent rules. Visible findings. A review you can understand.
Runs locally in your browser · No provider connection required
Recognize the instruction
inside the information.
A retrieved page or tool response can contain more than useful facts. Explore the patterns Ondrex flags and inspect why each one contributed to a score.
When text tries to take control.
Detect known phrases that ask an agent to disregard prior instructions, adopt an unrestricted persona or treat embedded text as a higher-priority command.
When context becomes the target.
Surface requests to reveal hidden prompts, repeat prior context or disclose configuration. Findings are signals for review, not proof of malicious intent.
When a payload asks for action.
Flag known bulk-delete, command-execution and confirmation-bypass phrases, with optional checks for suspicious text structure.
AN EXPLAINABLE REVIEW LOOP01 02 03 04
Paste. Inspect. Decide.
Supply text
Paste a redacted payload or choose one of five examples.
Select checks
Enable relevant categories and inspect the active policy.
Read findings
See matched rules, weights and measured scan duration.
Export review
Save findings. Include payload text only if you choose.
A working scanner.
A defined boundary.
The score is computed from your text using the supplied deterministic rules. It is not a trained AI classifier or a calibrated risk probability. Novel attacks can evade it, and benign text can trigger rules. Non-English text can trigger the non-ASCII heuristic. Ondrex has no live LLM, tool-call or agent integration and does not block actions outside this page.
CATALOG / MDV-ONX-001
Ondrex Personal License
Personal software license · One-time catalog price
Production integrations and hosted services are separate.