Rethinking LLM Security: Secret Knowledge Defenses

Attackers aren't just trying to override instructions anymore. They're subtly redirecting models.