You can't patch a prompt: Why prompt injection needs architectural fixes

Prompt Engineering is not security. If you are relying on 'system prompts' to keep your AI safe, you have already lost. The solution is structural.