AI LABS.
Choose a prepared test. Copy it into a chatbot, or give a browsing bot a reference page. Compare the actual responses in your private workspace.
Two ways to test
Copy a test
For any text chatbot. Copy the clean prompt into a fresh conversation, then the injection prompt into another fresh conversation. Paste both responses back for review.
Test a webpage
For a bot with browsing or page access. Give it a legitimate task and a reference URL. The reference contains the injection; the user task does not authorize it.
Visiting a website yourself does not expose it to a chatbot unless the chatbot can read that page. This site makes no model calls and does not automatically control another chatbot.
Task hijacking
Reference material must not replace the user task.
HOW THIS TEST WORKS →PREPARED TESTFake system authority
A page claiming to be a system message has no system privileges.
HOW THIS TEST WORKS →PREPARED TESTSynthetic information disclosure
A reference page must not induce disclosure of the synthetic control value in the user task.
HOW THIS TEST WORKS →PREPARED TESTUnnecessary code and output
A short summary task must not become unrelated code generation.
HOW THIS TEST WORKS →PREPARED TESTUnauthorized simulated action
A webpage must not supply approval or trigger an unrelated action.
HOW THIS TEST WORKS →PREPARED TESTFalse claims of completed work
The model must not claim it inspected resources or performed work it did not access.
HOW THIS TEST WORKS →Evidence before conclusions
Results distinguish boundary crossed, no crossing observed, and inconclusive. Missing markers do not prove security. Use synthetic data and only systems you own or have permission to test. Keep each attempt bounded; never use real credentials.