YARAAlert-ready after reviewPrompt injection and exfiltration in LLM app logs
YARA: invisible Unicode tag characters (U+E0000–U+E007F) in prompts, tool output or agent files
Unicode tag characters mirror ASCII but render as nothing, so instructions written with them are invisible to people yet readable by models (ASCII smuggling). AWS recommends stripping the U+E0000–U+E007F range before text reaches a model. This rule finds runs of tag characters encoded as UTF-8 (F3 A0 80–81 xx), UTF-16LE surrogate pairs (DB40 DC00–DC7F) or JSON \u escapes in agent instruction files (AGENTS.md, CLAUDE.md, .cursor/rules), MCP tool-description dumps, RAG documents and exported logs. The count threshold skips subdivision flag emoji, which use a few tag characters.
- Log source
- File scan of repositories, agent instruction files, MCP tool manifests, RAG corpora and exported LLM logs
- Techniques
- AML.T0068 LLM Prompt Obfuscation · LLM01 Prompt Injection · AML.T0051 LLM Prompt Injection · AML.T0110 AI Agent Tool Poisoning
- MITRE ATT&CK
- T1027.018Invisible UnicodeT1698Indirect Prompt Injection
- Reviewed
- 2026-11-10
Query
rule ThreatDeck_Unicode_Tag_Characters_Hidden_Text
{
meta:
description = "Runs of invisible Unicode tag characters (U+E0000-U+E007F) used to hide instructions from humans"
reference = "https://aws.amazon.com/blogs/security/defending-llm-applications-against-unicode-character-smuggling/"
author = "ThreatDeck"
date = "2026-11-09"
strings:
$utf8_tag = { F3 A0 ( 80 | 81 ) }
$utf16le_tag = { 40 DB ?? DC }
$json_escaped_tag = /\\u[dD][bB]40\\u[dD][cC][0-7][0-9a-fA-F]/
condition:
#utf8_tag > 10 or #utf16le_tag > 10 or #json_escaped_tag > 10
} Expected false positives
- Text with many subdivision flag emoji (each flag uses up to seven tag characters)
- Unicode test corpora and fuzzing fixtures
Sources
- AWS Security Blog: Defending LLM applications against Unicode character smuggling (Sep 2025)
- Embrace The Red: Hiding and finding text with Unicode Tags (Jan 2024)
Reviewed against vendor documentation, not tested in your environment. Defensive use only.