Workshop · trust boundary
Try to break the classifier
Pick a payload, toggle defences, and compare what the model was shown. Escaping is structural; everything after that is judgement — which is why money decisions are checked by arithmetic.
The baseline attack: a plain instruction to override the classification. If this one works, nothing else matters.
Pick a payload and hit Classify. Run it once with the defences on and once with them off — the difference is easiest to see on
inj-02, which closes the delimiter and opens a forged system block.