Definition
Reverse prompt engineering is the attempt to reconstruct an unknown prompt from outputs produced under that prompt. An observer may compare outputs to infer candidate wording and test whether a candidate prompt produces similar responses.
The result may be an exact reconstruction, a partial reconstruction, or only a functionally similar approximation. Similar outputs do not prove that the inferred wording matches the original prompt. Asking a model to disclose hidden instructions is prompt extraction. Inferring routing implemented in application code is broader system reverse engineering.
Simple example
A support assistant answers disputed refund requests by asking for an order number and saying a human will review the case. A user tries several phrasings, then writes a candidate prompt that reproduces the observed pattern. The candidate may mimic the behavior without matching the original wording. Application code might also control the format.
If the application enforces authorization outside the model, reconstructing the prompt does not bypass those controls.
Why it matters
Prompts can contain internal workflow labels or operational details that an application owner would prefer not to expose. Repeated observation can also help someone craft a more targeted prompt injection attempt.
Teams should keep sensitive data out of prompts and assume that observable behavior can reveal aspects of hidden instructions or application logic.
One important nuance
Keeping a system prompt hidden is not a robust security control. Prompt text can leak or be approximated from behavior, and the model may not follow it. Never place credentials in prompts or rely on secret wording to enforce permissions. As OWASP recommends, enforce authorization outside the model. Validate actions and keep an audit trail even if every instruction becomes public.