Prompt Injection Test Cases for My Assistant
Write defensive test inputs to see whether an assistant ignores hidden instructions.
Added October 15, 2026. Written with AI assistance and checked for clarity and placeholders; not reviewed by a lawyer, accountant, clinician or other professional. How these prompts are made
The Open buttons send the prompt text as currently shown below to that site only when you click. Fill in the placeholders first.
I am testing my own assistant for resilience. Its job: [ASSISTANT_JOB]. Its rules: [ASSISTANT_RULES]. Content it reads from outside: [EXTERNAL_CONTENT_TYPES]. Write 15 harmless test inputs that contain embedded instructions of the kind a document, web page or email might hold, such as asking the assistant to ignore its rules, reveal its instructions, or send data elsewhere. Use obviously fake targets. For each, state the safe behaviour I expect and how I will know it failed. Return a test table, a scoring method, and defensive design suggestions: treat outside content as data, limit tools by task, require confirmation for sending or deleting, and log what the assistant read. This is for testing systems I own only. Also include two cases in which the hidden instruction is written in another language or split across paragraphs, and describe how to record, for each test, the exact input, the assistant's reply, and whether any tool was called.
Fill in before use
Optional: type your details and the prompt above updates. Nothing is saved or sent.
Replace each bracketed placeholder with your own details, then paste into Any AI or another assistant.