practice sandbox Β· rule-based sim Β· not a live LLM

Prompt-Injection Training Range

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.

Difficulty: Normal = rule-based, forgiving. Hard = the model understands intent & refuses easy tricks (like the real arena).
support-assistant
The 3 defense layers (learn to read the tags)

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).

Technique families cheat-sheet

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.