Hone extraction & evasion on a target that tells you which defense layer blocked you. Diagnose β pick a technique β combine. Then take the skill to the real arena β with your own hands.
INPUT β a keyword filter on YOUR text (before the model sees it). Beat it with paraphrase / synonyms / obliqueness.
MODEL β the assistant itself refusing (it understands intent). Beat it by reframing, or making the leak a byproduct of a legit task.
OUTPUT β a filter on the RESPONSE that catches secret-shaped values. Beat it by reshaping the output (split / space / reverse / encode).
Prompt-leak: make it repeat/echo its own context. Β· Task-byproduct: translate/summarize/continue a doc so content leaks incidentally. Β· Roleplay: reframe as test/debug/dummy. Β· Format: ask for JSON/table/list. Β· Obfuscation: spaced/reversed/base64 to dodge output filter. Β· Compound: combine β usually required.