How hidden prompts work
Unicode has a block of tag characters (U+E0000 to U+E007F) that mirror every ASCII letter but render as nothing at all. Browsers, editors, chat apps and email clients show them as blank. Many AI models, though, read them like normal text. An attacker can hide a sentence like “ignore previous instructions and …” inside an innocent message, a technique known as ASCII smuggling.
A related trick hides bytes inside runs of variation selectors attached to an emoji, so a single 😀 can carry a whole message.
These techniques have been used in phishing campaigns to slip past keyword filters, and in prompt-injection attacks against AI assistants and agents that read emails, web pages and documents.
What ClepBoard does
- Detects tag characters, stray variation selectors, zero-width characters and direction overrides.
- Decodes hidden tag text and variation-selector payloads and shows you the actual message in a red warning.
- Removes them so you can copy a safe version. Legitimate uses, like the tag characters inside the 🏴 England and Scotland flags, are preserved.
When to use it
- Before pasting an email, PDF, job application or web page into ChatGPT, Claude or Copilot
- When reviewing documents that will be fed to an AI agent or RAG pipeline
- When a model behaves strangely after reading a piece of text
- When checking user-submitted content for smuggled instructions