Prompt injection stems from a fundamental characteristic of current LLM architectures, rather than a vulnerability that can simply be patched.