How One Developer Built Invisible-Character Defenses Against LLM Prompt Injection

A developer running an AI-powered travel site identified a security risk where hidden Unicode characters — invisible on screen but readable by language models — could be used to smuggle malicious instructions into prompts. To counter this, they built a sanitization function that strips Unicode tag blocks, zero-width characters, bidirectional controls, and repeated character runs before any user text reaches an AI model. A separate wrapping function encloses user input within randomly generated boundary tags on every request, making it far harder for attackers to escape the data context and issue direct commands. The system also includes explicit prompt-level rules instructing the model to treat all fenced content as data, never as instructions. The developer notes that no single layer is sufficient on its own, and that defense requires combining input sanitization, structural wrapping, and prompt constraints together.