Audit a prompt for injection risk
Reviews your own system prompt the way an attacker would read it, and points to the exact sentence that gives something away.
Tested on Claude Sonnet · 4.5 ·
AI Place Hub
39
You are reviewing a system prompt for injection resistance. Here it is:
{{system_prompt}}
The application inserts untrusted content here: {{where_untrusted_content_goes}}
The model has access to these tools: {{tools}}
Report, in this order:
1. Every place where the prompt does not distinguish instructions from data. Quote the sentence.
2. What an attacker could achieve with the listed tools if the model followed an injected instruction. Be concrete about the worst realistic outcome, not the most dramatic one.
3. Any sentence that leaks information about the system: internal names, other users, the model in use, the shape of the data.
4. The three changes with the best effect for the least text, in the order you would make them.
Do not rewrite the prompt. Do not say it is generally secure. If a defence only works when the input is well behaved, say so.Notes from the author
Point two is where it pays off. Most prompt reviews list vague weaknesses. Make the model say what an attacker gets, given the tools you listed, and the review ends in a decision about which tools to remove.
For information only, not professional advice. You use it at your own risk and we are not liable for any loss or damage. See the disclaimer.
Tagged