Prompt injection exploits the fact that language models do not reliably distinguish instructions from content. Text embedded in a document, email, or webpage can be crafted to instruct the model to ignore its directions, reveal information, or produce misleading output.
Indirect injection is the more serious variant in legal settings, since it arrives through documents the system was asked to analyze rather than through user input.
Alternative Names:
Indirect Prompt Injection, Injection Attack
Why it Matters?
Litigation systems process adversarial material by design: opposing party productions, third-party documents, and unvetted records. An instruction hidden in a produced document could in principle alter how a summarization or extraction system reports on it. This is an emerging rather than a common problem, and the practical mitigations are content and instruction separation, output validation, and attorney review of anything consequential.
Frequently Confused with
Related terms
Frequently asked questions
Why is prompt injection a concern in litigation tools?
How is it mitigated?
