Multiple Choice Flashcards

Contain risky responses, preserve prompts and logs, test injection paths, enforce retrieval permissions, and update filters Increase the model temperature so responses become less predictable Remove all logging to prevent further exposure of sensitive prompts Treat the issue as user training only because the model followed a customer instruction