adapters

Live demo of a prompt-injection defense for LLMs using per-position semantic overlays. The page shows how marked spans can be treated as untrusted while the base model stays frozen, with interactive examples for retrieved text, hidden instructions, language marking, and request-scoped overlays.