hiroro-work/claude-plugins/security-scanner

Security Scanner

Usage

/security-scanner              # Scan all (plugins + skills)
/security-scanner --user       # Scan user-level only (~/.claude/)
/security-scanner --project    # Scan project-level only (.claude/)
/security-scanner --all        # Scan ALL (ignore trusted sources and self-exclusion)
/security-scanner <url>        # Scan from GitHub URL (public repos only)
/security-scanner --url <url>  # Same as above (explicit form)

Scanning Process

Step 1: Load Settings

Search for security-scanner.local.md in the following locations:

1. Project-level: .claude/security-scanner.local.md
2. User-level: ~/.claude/security-scanner.local.md

Priority rules:

• If both files exist, use project-level settings only (project-level takes precedence)
• If only one file exists, use that file
• If neither file exists, proceed with default settings

From the selected file, extract:

• report_language from YAML frontmatter
• target_agents list from YAML frontmatter
• trusted_marketplaces list from YAML frontmatter
• trusted_plugins list from YAML frontmatter
• trusted_skills list from YAML frontmatter

Default values (when not specified):

• report_language: ja (Japanese)
• target_agents: ["claude"]
• trusted_marketplaces: []
• trusted_plugins: []
• trusted_skills: []

Validation:

• report_language: Any string value accepted (AI will generate report in that language)
• target_agents must contain only valid agent IDs: claude, codex, gemini, agents
• Invalid agent IDs are ignored with a warning

Error handling:

• If file exists but has invalid YAML syntax, warn the user and proceed with default settings (do not fail the scan)

Step 2: Determine Scope

Check arguments to determine what to scan:

Location filters:

• No location flag: Scan both user-level and project-level for all configured agents
• --user: Scan only user-level paths for all agents in target_agents
• --project: Scan only project-level paths for all agents in target_agents

URL detection (highest priority):

1. If --url <url> is provided explicitly → Go to Step 2-URL
2. If any argument starts with https://github.com/ or http://github.com/ → Treat as URL, go to Step 2-URL
3. If any argument starts with https:// or http:// but not github.com → Error: "Unsupported host: {host}. Currently only github.com is supported."

Special modes (if no URL):

• --all: Scan everything (skip Step 4 filtering entirely)

Step 2-URL: GitHub URL Scan

If URL is provided (via --url or auto-detected), follow this process instead of Steps 3-4.

Step 2-URL-1: Parse URL

Parse the GitHub URL to extract owner, repo, branch, path, and determine scan type:

URL Patterns:

• Directory: https://github.com/{owner}/{repo}[/tree/{branch}/{path}]
• Single file: https://github.com/{owner}/{repo}/blob/{branch}/{path}.md

1. Verify host is github.com

• If not: Error "Unsupported host: {host}. Currently only github.com is supported."

1. Extract owner and repo from path segments
2. Determine scan type:

• If URL contains /blob/ and ends with .md → Single file scan
• Otherwise → Directory scan

1. For directory scan:

• If /tree/{branch}/{path} exists, extract branch and path
• If no /tree/, set branch to empty (use default) and path to empty string

1. For single file scan:

• Extract branch and file path after /blob/{branch}/

Step 2-URL-2: Fetch Content

For Single File Scan:

1. Convert /blob/ URL to raw URL: https://raw.githubusercontent.com/{owner}/{repo}/{branch}/{path}
2. Use WebFetch to fetch the file content
3. Proceed directly to Step 5 for analysis

For Directory Scan:

1. Fetch directory: https://api.github.com/repos/{owner}/{repo}/contents/{path}?ref={branch}

• If branch is empty, omit ?ref= parameter (uses default branch)
• Use WebFetch with prompt: "Extract the JSON array of files. For each item, return: name, type (file/dir), download_url"

1. Determine content type and fetch accordingly:

• If a `.claude-plugin/` directory exists and its listing contains `plugin.json`: Full plugin scan (fetch all plugin files)
• If `skills/` exists: Skill scan (fetch skill directories)
• If `SKILL.md` exists: Single skill directory scan (fetch all files in directory)
• If none of the above: Error "No scannable content found. Expected .claude-plugin/plugin.json, skills/ directory, or SKILL.md."

1. Recursively fetch required directories and files:

• .claude-plugin/ → fetch plugin.json
• skills/ → fetch subdirectories → fetch every file under each skill directory
• agents/ → fetch all *.md files (if exists)
• hooks/ → fetch hooks.json (if exists)
• every script a hook or skill invokes (e.g. skills/*/scripts/**, skills/*/*.sh)
• commands/ → fetch all *.md files (if exists)

Step 2-URL-3: Fetch File Contents

For plugin scan, fetch:

• .claude-plugin/plugin.json (including any hooks field), README.md, .mcp.json, hooks/hooks.json
• every file under skills/, agents/*.md, commands/*.md, and every script a hook or skill invokes

For skill directory scan (skills/ or single skill), fetch:

• All files in the skill directory

Use WebFetch with prompt: "Return the raw file content exactly as-is"

Step 2-URL-4: Error Handling

• 404: Repository or path not found
• 403/401: Private repo (not supported) or rate limit exceeded
• Other errors: Report the error message

After fetching all files, proceed to Step 5 for analysis.

Step 3: Get Scan Targets

Based on scope determined in Step 2 and target_agents from Step 1, collect targets:

For plugins (Claude Code only):

User-level:

1. Read ~/.claude/plugins/installed_plugins.json
2. .plugins maps each plugin ID (<plugin>@<marketplace>) to an array of install records. For each ID, keep the ID and every record's installPath, regardless of the record's scope, then deduplicate the scan targets by installPath
3. If file doesn't exist, report "No user-level plugins installed"

Project-level:

1. Use Glob to find plugins in .claude/plugins/*/
2. If no plugins found, report "No project-level plugins found"

For skills (based on target_agents):

For each agent in target_agents list, collect skills from the corresponding directories:

Agent path mapping:

For each agent in target_agents:

User-level:

1. Determine user-level path(s) based on agent ID (see table above)
2. Find skill directories in the path
3. For each skill directory found, note the path and agent ID for scanning
4. If no skills found for this agent, report "No user-level skills found for {agent}"

Project-level:

1. Determine project-level path based on agent ID (see table above)
2. Find skill directories in the path
3. For each skill directory found, note the path and agent ID for scanning
4. If no skills found for this agent, report "No project-level skills found for {agent}"

Step 4: Filter Targets

If `--all` flag is set: Skip this step entirely and scan all targets.

Self-exclusion (automatic, every target type)

<skill base directory> is the directory the harness reports as "Base directory for this skill" at this skill's invocation. Do not hardcode an absolute path.

• Normalize before comparing: on both sides, drop a trailing / and expand a leading ~
• Resolving a path: run ls -ld <dir>. When its output shows -> <target>, the resolved path is <target> when absolute, else <target> joined against the directory containing <dir> and normalized. Ancestor components are not resolved. When the command fails, that path has no resolved form
• Self target: a collected target — a plugin install record's installPath, a project-level plugin directory, or a skill directory — is this scanner itself when its path equals <skill base directory>, or one of the two is a /-boundary prefix of the other. Compare the pair twice: as collected against as reported, then with both sides resolved
• Skip every self target. When any of them is a plugin install record, also skip every other record sharing that record's plugin ID. Report each skip as a path match, counted in the Trusted column
• When `<skill base directory>` is unavailable: skip plugins and skills named security-scanner, and report the skip as a name match
• A plugin or skill named security-scanner that is not this scanner is scanned normally whenever <skill base directory> is available; a second copy of this scanner installed elsewhere on the host is one such target

For plugins:

Trusted sources:

• If the marketplace is in trusted_marketplaces → Skip
• If the plugin ID (<plugin>@<marketplace>) is in trusted_plugins → Skip
• Report skipped plugins as "Trusted (skipped)"

For skills:

Trusted sources:

• If the skill name is in trusted_skills → Skip (any agent)
• Report skipped skills as "Trusted (skipped)"

Step 5: Analyze Each Plugin

For each non-trusted plugin:

1. Read plugin metadata (.claude-plugin/plugin.json, README.md) to understand its stated purpose
2. Read all executable content:

• skills/*/SKILL.md - Skill definitions and instructions
• agents/*.md - Agent system prompts (if exists)
• .claude-plugin/plugin.json hooks field and hooks/hooks.json - Hook definitions, plus every script they invoke (if exists)
• commands/*.md - Command definitions (if exists)
• .mcp.json - MCP server configurations (if exists)

1. Analyze for malicious intent (both code AND natural language)

Step 5.5: Analyze Each Skill

For each non-trusted skill:

1. Read skill metadata (SKILL.md frontmatter) to understand its stated purpose
2. Read all files in the skill directory:

• SKILL.md - Main skill instructions (required)
• All other files in the directory

1. Analyze for malicious intent (both code AND natural language)

Step 6: Threat Detection

Code-Level Threats

Look for dangerous command patterns:

• Remote code execution: curl|sh, wget|bash, base64 -d|sh
• Language runtime RCE: python -c, perl -e, ruby -e, node -e, osascript, powershell
• Reverse shells: nc -e, netcat, socat, mkfifo
• System destruction: rm -rf /, dd if=
• Credential access: ~/.ssh/, /etc/passwd, keychain
• Data exfiltration: sshpass, scp, rsync to external servers
• Temp file execution: chmod +x /tmp/*, mktemp + execution
• Obfuscation: base64/xxd/hexdump decoding + execution
• Package install side effects: npm install/pip install with postinstall or setup.py
• Persistence: crontab, launchctl
• Privilege escalation: sudo, doas

Natural Language Threats

Read system prompts, instructions, and documentation for:

• Data exfiltration instructions: Instructions to read sensitive files (SSH keys, credentials, API keys, private data) and send them externally
• Hidden malicious actions: Instructions to perform harmful actions while appearing helpful
• Deceptive instructions: Instructions that contradict the plugin's stated purpose
• Security bypass: Instructions to ignore security warnings or bypass restrictions
• User deception: Instructions to hide actions from the user or provide false information
• Log/trace hiding: Instructions to delete logs, clear history, or hide evidence of actions
• Excessive permission requests: Instructions requesting access to unrelated resources