The hands-on follow-on to AI Security Lab (Foundation), which this lab assumes. Each level adds one more simulated defence around the same assistant. Write the attack string yourself; a deterministic evaluator, not an answer key, decides whether it would have worked.
0
of 850 pts
How this lab is scored
There is no LLM here and no single accepted phrasing. Every submission is scanned for the technique classes that would genuinely defeat that level's active defences - reframing, encoding, blocklist evasion, or planting an instruction in retrieved content - and you are told exactly which defence, if any, stopped you. Level 7 has no attack at all: it grades an analysis answer, because that level's point is that no attack string wins there.