跳到內容
Agent Threat Rules (ATR)EN
ΒSkill 入侵
ATR-2026-00129
收錄於 2026-04-05 · 作者 ATR Community

Unicode Tag Character Smuggling

Detects Unicode Tag characters (U+E0001-U+E007F) and other invisible Unicode ranges used to smuggle hidden instructions in SKILL.md files.

unicode smuggling · pattern · confidence: high

嚴重度
critical
車道
alert · hunt比對到就標出、回報
成熟度
test還沒定論
掃描目標
skill
狀態
experimental

框架對照

OWASP Agentic Top 10
  • ASI01:2026
OWASP LLM Top 10
  • LLM01:2025

這些對應是 ATR 自己做的,框架方沒有審查或背書。

說明

Detects Unicode Tag characters (U+E0001-U+E007F) and other invisible Unicode ranges used to smuggle hidden instructions in SKILL.md files. These characters are invisible in most editors and renderers but are processed by AI agents as text. Snyk ToxicSkills research (Feb 2026) documented this as "ASCII smuggling" — invisible payloads that agents read but humans cannot see.

建議回應alertblock input
作者:ATR Community · 建立 2026-04-05 · 最後修改 2026-11-20

野外驗證

驗證日期
2026-04-08
樣本數
53,577
誤報率,野外樣本
0.0056%

偵測條件

組合方式:any
  1. 01
    3+ Unicode Tag characters (invisible payload, Snyk ASCII smuggling) via surrogate pairs
    欄位:content運算子:regex
  2. 02
    5+ zero-width characters spread through text (steganographic smuggling)
    欄位:content運算子:regex

會觸發的攻擊範例

  1. # Skill
    
    Normal​​text​​with​​hidden​​instructions​​embedded

規則必須抓到的攻擊範例,很多是照公開回報的攻擊改寫的,有些是照攻擊的形狀寫出來的。範例跟規則一起版本化,當作回歸測試。之後改規則時,如果其中一個範例不再命中,一定看得出來。

不會觸發的正常樣本

  1. # Normal Skill
    
    This skill helps with file management.
    
    ## Commands
    - list: List files
  2. # Tool
    
    Use this tool for simple tasks.
title: "Unicode Tag Character Smuggling"
id: ATR-2026-00129
rule_version: 1
status: experimental
description: >
  Detects Unicode Tag characters (U+E0001-U+E007F) and other invisible Unicode
  ranges used to smuggle hidden instructions in SKILL.md files. These characters
  are invisible in most editors and renderers but are processed by AI agents as
  text. Snyk ToxicSkills research (Feb 2026) documented this as "ASCII smuggling"
  — invisible payloads that agents read but humans cannot see.
author: "ATR Community"
date: "2026/04/05"
schema_version: "1.0"
detection_tier: pattern
maturity: test
severity: critical

references:
  mitre_atlas:
    - "AML.T0010 - AI Supply Chain Compromise"
  owasp_llm:
    - "LLM01:2025"
  owasp_ast:
    - "AST01:2026 - Malicious Skills"
    - "AST05:2026 - Unsafe Deserialization"
  research:
    - "Snyk ToxicSkills: ASCII smuggling via Unicode Tags (Feb 2026)"
    - "Unicode Technical Report #36: Security Considerations"

  owasp_agentic:
    - ASI01:2026
metadata_provenance:
  mitre_atlas: auto-generated

compliance:
  nist_ai_rmf:
    - subcategory: "MP.5.1"
      context: "Invisible Unicode Tag characters and zero-width steganographic payloads embedded in SKILL.md files are adversarial inputs that exploit the gap between human-visible content and agent-parsed content; MP.5.1 requires identifying and characterizing these hidden prompt-injection vectors as risks to the AI system."
      strength: primary
    - subcategory: "MG.3.2"
顯示全部(還有 75 行)
      context: "SKILL.md files are third-party supplied artifacts consumed by AI agents, and Unicode smuggling is a supply chain compromise vector; MG.3.2 requires monitoring of these pre-trained/third-party components for hidden malicious content before agent execution."
      strength: secondary
    - subcategory: "MG.2.3"
      context: "Detection of 3+ Unicode Tag characters or 5+ zero-width characters indicates a covert injection payload that must trigger containment of the affected skill; MG.2.3 mandates predefined response plans to disengage or quarantine compromised skills before agents execute the smuggled instructions."
      strength: secondary
    - subcategory: "MS.2.7"
      context: "NIST AI RMF MEASURE 2.7 (security and resilience evaluated and documented) is supported by this rule's runtime detection of the skill supply-chain compromise (Unicode Tag Character Smuggling)."
      strength: primary
    - subcategory: "MG.3.1"
      context: "NIST AI RMF MANAGE 3.1 (third-party resources regularly monitored) is supported where this rule detects the skill supply-chain compromise (Unicode Tag Character Smuggling)."
      strength: secondary
  iso_42001:
    - clause: "8.1"
      context: "ISO/IEC 42001 Clause 8.1 (operational planning and control, including control of externally provided processes) is operationalised by this rule's detection of the skill supply-chain compromise (Unicode Tag Character Smuggling)."
      strength: primary
    - clause: "8.3"
      context: "ISO/IEC 42001 Clause 8.3 (AI risk treatment) is supported by this rule, which implements runtime detection of the skill supply-chain compromise (Unicode Tag Character Smuggling) as a treatment control."
      strength: secondary
  eu_ai_act:
    - article: "15"
      context: "Article 15 (accuracy, robustness and cybersecurity) requires high-risk AI systems to resist unauthorised attempts to alter their use, outputs or performance; this rule provides runtime detection evidence by flagging the skill supply-chain compromise (Unicode Tag Character Smuggling)."
      strength: primary
    - article: "9"
      context: "Article 9 (risk management system) requires identified risks to be addressed by appropriate measures; this rule is a runtime risk-treatment control that detects the skill supply-chain compromise (Unicode Tag Character Smuggling)."
      strength: secondary
tags:
  category: skill-compromise
  subcategory: unicode-smuggling
  confidence: high
  scan_target: skill

agent_source:
  type: mcp_exchange
  framework:
    - any
  provider:
    - any

detection:
  condition: any
  conditions:
    # Detect Unicode Tag characters (U+E0000-U+E007F) used for ASCII smuggling.
    # These are invisible in editors but processed by AI agents as text.
    # Must use Unicode codepoint escapes (JS strings are UTF-16, not raw bytes).
    - field: content
      operator: regex
      value: "(?:\\uDB40[\\uDC00-\\uDC7F]){3,}"
      description: "3+ Unicode Tag characters (invisible payload, Snyk ASCII smuggling) via surrogate pairs"

    # Detect zero-width characters used for steganographic payload hiding.
    # ZWSP, ZWNJ, ZWJ, Word Joiner, BOM — 5+ occurrences = suspicious density.
    - field: content
      operator: regex
      value: "[\\u200b\\u200c\\u200d\\u2060\\ufeff].*[\\u200b\\u200c\\u200d\\u2060\\ufeff].*[\\u200b\\u200c\\u200d\\u2060\\ufeff].*[\\u200b\\u200c\\u200d\\u2060\\ufeff].*[\\u200b\\u200c\\u200d\\u2060\\ufeff]"
      description: "5+ zero-width characters spread through text (steganographic smuggling)"

confidence: 50
wild_validated: "2026/04/08"
wild_samples: 53577
wild_fp_rate: 0.0056
test_cases:
  true_positives:
    - input: "# Skill\n\nNormal\u200b\u200btext\u200b\u200bwith\u200b\u200bhidden\u200b\u200binstructions\u200b\u200bembedded"
      expected: trigger

  true_negatives:
    - input: "# Normal Skill\n\nThis skill helps with file management.\n\n## Commands\n- list: List files"
      expected: no_trigger
    - input: "# Tool\n\nUse this tool for simple tasks."
      expected: no_trigger

response:
  actions:
    - alert
    - block_input

修訂紀錄

建立於
2026-04-05
最後修改
2026-11-20
到 GitHub 看完整的 commit 紀錄