Invisible and reordered text: bidirectional controls, tag characters and confusables in code and prompts

Goal

Make text that looks harmless in a review tool or chat window reveal the characters it actually contains, before it is merged or passed to a model.

Prerequisites

Access to the raw text (diff, prompt, fetched document) as code points, not only its rendering.

Steps

1. Know the three families:
   • Bidirectional controls (for example the embedding, override and isolate characters U+202A–U+202E and U+2066–U+2069). The Trojan Source research showed that they can reorder source code on screen so that a reviewer sees different logic from what the compiler processes.
   • Invisible characters: zero-width space and joiners, and the Unicode Tags block (U+E0000–U+E007F), whose characters mirror ASCII but render as nothing in most interfaces. Security researchers have shown that text written in tag characters can carry instructions to a model while staying invisible to the person who pastes it.
   • Confusables: characters from other scripts that look like Latin letters. Unicode Technical Standard #39 defines confusable detection and mixed-script checks for exactly this.
2. Add a CI check that fails on bidirectional control characters in source files unless a file is explicitly allowlisted (for example translation resources that need them).
3. Before passing fetched or pasted text to a model, strip or flag tag characters and zero-width characters, and log that you did.
4. For identifiers, package names and domains, apply a mixed-script or confusable check following UTS #39 rather than a hand-written list.
5. In review tools, enable the option to show hidden characters; many editors and code hosts now warn on bidirectional text.
6. When reporting a finding, show code points (U+202E), never the raw characters, so the report itself cannot mislead.

Expected result

Hidden or reordering characters are caught mechanically; human review sees an explicit marker instead of misleading text.

Limits and test basis

Legitimate right-to-left text and emoji sequences use some of these characters; blanket removal can corrupt content, so scope the check to code and to text headed for a model. Rendering differs between tools, so "it looks fine here" proves nothing.