Prompt injection and defenses

Prompt injection is the adversarial manipulation of an LLM's behavior by embedding malicious instructions in untrusted input. It is the most critical security vulnerability in production LLM systems.