ACB-GEN-004: Artifacts contain no invisible or bidirectional control characters
| Field | Value |
|---|---|
| Status | draft |
| Severity | high |
| Check type | static |
| Applies to | skill, instruction-file, mcp-server-declaration |
| Profiles | strict-general |
| Exceptions | not-permitted |
Description
No artifact contains Unicode characters that are invisible or that reorder displayed text, so that what a human reviewer sees is what the model reads.
Rationale
Models read code points, reviewers read rendered glyphs. Zero-width characters, Unicode tag characters and bidirectional overrides let an author hide instructions that render as nothing, or as something else, in diffs, editors and registry pages. This is a cheap, reliable way to get unreviewed instructions into model context.
Audit procedures
forbidden-code-points (pattern)
No file contains any of: U+200B-U+200F, U+202A-U+202E, U+2060-U+2064, U+2066-U+2069, U+FEFF (except as the first code point of a file), U+00AD, U+180E, U+034F, U+061C, U+3164, U+2800, U+FFF9, U+FE00-U+FE0F (variation selectors), U+E0100-U+E01EF (supplementary variation selectors), any code point in U+E0000-U+E007F (tag characters), or any code point in Unicode private use areas.
confusable-identifiers (pattern)
Skill names, file names and command names in code blocks contain only ASCII, so mixed-script look-alikes cannot impersonate a trusted name.
Remediation
Strip the characters, review the decoded content to determine what was hidden, and treat the artifact as potentially malicious until the origin of the characters is explained.
References
- OWASP Top 10 for Agentic Applications 2026: ASI01
- Careful Adoption of Agentic AI Services risk areas: behaviour, accountability
- NIST CSF 2.0: ID.RA-09
- https://www.unicode.org/reports/tr36/
- https://trojansource.codes/