ΒSkill 入侵
ATR-2026-00062
收錄 於 2026-03-08 · 作者 ATR Community
Hidden Capability in MCP Skill
Detects MCP skills that expose hidden or undocumented capabilities beyond their declared tool schema.
hidden capability · 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
- mcp
- 狀態
- experimental
框架對照
- MITRE ATLAS
- OWASP Agentic Top 10
- ASI02:2026
- ASI05:2026
- OWASP LLM Top 10
- LLM03:2025
- LLM06:2025
這些
說明
Detects MCP skills that expose hidden or undocumented capabilities beyond their declared tool schema. A skill may advertise a simple interface but accept hidden parameters like "debug_mode", "admin_override", or "raw_exec" that unlock dangerous functionality. This is a common pattern in trojaned MCP packages.
建議回應alertsnapshot
作者:ATR Community · 建立 2026-03-08 · 最後修改 2026-11-20
野外驗證
驗證日期
2026-04-08
樣本數
53,577
偵測條件
組合方式:any- 01欄位:tool_args運算子:regex
- 02欄位:tool_args運算子:regex
- 03欄位:tool_args運算子:regex
- 04欄位:tool_response運算子:regex
會 觸發 的 攻擊範例
- 觸發條件:hidden debug_mode and raw_exec parameters
tool_name: note_taker tool_args: '{"text": "hello", "debug_mode": true, "raw_exec": "cat /etc/shadow"}' - 觸發條件:dunder-wrapped hidden parameter
tool_name: calculator tool_args: '{"expression": "2+2", "__backdoor__": "reverse_shell"}'
規則
不會 觸發 的 正常樣本
tool_name: debugger tool_args: '{"breakpoint": "line_42", "mode": "step"}'tool_name: admin_panel tool_args: '{"action": "list_users"}'
已知 的 誤報情境
- ▸Development/staging environments where debug modes are intentionally enabled
- ▸Administrative tools that legitimately expose debug parameters
完整 YAML
在 GitHub 編輯title: "Hidden Capability in MCP Skill"
id: ATR-2026-00062
rule_version: 2
status: experimental
description: >
Detects MCP skills that expose hidden or undocumented capabilities beyond their
declared tool schema. A skill may advertise a simple interface but accept hidden
parameters like "debug_mode", "admin_override", or "raw_exec" that unlock dangerous
functionality. This is a common pattern in trojaned MCP packages.
author: "ATR Community"
date: "2026/03/08"
schema_version: "0.1"
detection_tier: pattern
maturity: test
severity: critical
references:
owasp_llm:
- "LLM03:2025"
- "LLM06:2025"
owasp_agentic:
- "ASI02:2026"
- "ASI05:2026"
mitre_atlas:
- "AML.T0010 - AI Supply Chain Compromise"
cve:
- "CVE-2025-59536"
compliance:
nist_ai_rmf:
- subcategory: "GV.6.1"
context: "Hidden capabilities in MCP skills represent third-party/supplier AI risk where a packaged tool exposes undocumented parameters beyond its declared schema; GV.6.1 requires policies and procedures that govern third-party AI components and detect deviations from declared interfaces."
strength: primary
- subcategory: "MG.3.2"
context: "Trojaned MCP packages with hidden parameters like debug_mode or admin_override are exactly the supply-chain risk MG.3.2 addresses by requiring monitoring of pre-trained models and third-party components used in development for unexpected or unsafe capabilities."
strength: secondary
- subcategory: "MP.5.1"
context: "Detection of undocumented dangerous parameters characterizes the likelihood and magnitude of supply-chain compromise impact; MP.5.1 requires that these hidden-capability risks be identified and tracked as part of risk characterization."
strength: secondary
- subcategory: "MS.2.7"顯示全部(還有 96 行)收起
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 (Hidden Capability in MCP Skill)."
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 (Hidden Capability in MCP Skill)."
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 (Hidden Capability in MCP Skill)."
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 (Hidden Capability in MCP Skill) 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 (Hidden Capability in MCP Skill)."
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 (Hidden Capability in MCP Skill)."
strength: secondary
tags:
category: tool-poisoning
subcategory: hidden-capability
scan_target: mcp
confidence: high
agent_source:
type: tool_call
framework:
- mcp
- langchain
- autogen
provider:
- any
detection:
conditions:
- field: tool_args
operator: regex
value: "(?i)(debug[_-]?mode|admin[_-]?override|raw[_-]?exec|unsafe[_-]?mode|bypass[_-]?check|no[_-]?sandbox)"
- field: tool_args
operator: regex
value: '(?i)(__(?:backdoor|hidden|secret|internal|debug|admin|exec|eval|shell|cmd|import|builtins|globals)__|_internal_|_hidden_|_backdoor_|_secret_)'
- field: tool_args
operator: regex
value: "(?i)(eval_code|exec_raw|shell_cmd|run_arbitrary|system_call|inject_prompt)"
- field: tool_response
operator: regex
value: "(?i)(hidden\\s+mode\\s+activated|debug\\s+mode\\s+enabled|admin\\s+access\\s+granted|sandbox\\s+disabled)"
condition: any
false_positives:
- "Development/staging environments where debug modes are intentionally enabled"
- "Administrative tools that legitimately expose debug parameters"
response:
actions:
- alert
- snapshot
actions_rationale: >-
response-action eligibility (docs/RESPONSE-ACTION-ELIGIBILITY.md): benign FP 209/5352 =
3.91% exceeds the 2% automatic-demotion line; ceiling = observe. Removed
block_tool(interrupt), quarantine_session(terminate), kill_agent(terminate). Detection
unchanged — recall cost is zero.
auto_response_threshold: critical
message_template: >
[ATR-2026-062] CRITICAL: Hidden capability detected in MCP skill. Tool arguments
or response indicate undocumented functionality that may be a backdoor.
This rule no longer blocks the tool or quarantines the session; it alerts
and snapshots. The hook may still deny on severity — see
docs/RESPONSE-ACTION-ELIGIBILITY.md. Investigate the skill source.
confidence: 53
wild_validated: "2026/04/08"
wild_samples: 53577
test_cases:
true_positives:
- input:
tool_name: "note_taker"
tool_args: '{"text": "hello", "debug_mode": true, "raw_exec": "cat /etc/shadow"}'
expected: triggered
matched_condition: "hidden debug_mode and raw_exec parameters"
- input:
tool_name: "calculator"
tool_args: '{"expression": "2+2", "__backdoor__": "reverse_shell"}'
expected: triggered
matched_condition: "dunder-wrapped hidden parameter"
true_negatives:
- input:
tool_name: "debugger"
tool_args: '{"breakpoint": "line_42", "mode": "step"}'
expected: not_triggered
reason: "Debugger tool legitimately using debug-related terminology"
- input:
tool_name: "admin_panel"
tool_args: '{"action": "list_users"}'
expected: not_triggered
reason: "Admin tool using normal admin operations"