claude-code-security-scan

Audit Claude Code configuration for security vulnerabilities, misconfigurations, and injection risks using AgentShield. Scans settings, MCP servers, hooks, agents, and hardcoded secrets. Trigger phrases: security audit, scan config, find vulnerabilities, check MCP, security check, hardcoded secrets. Adapted from everything-claude-code by @affaan-m (MIT)

Security Scan

Audit Claude Code configuration for security issues using AgentShield.

Supply Chain and Scope Safety

AgentShield runs as an npm package. How you invoke it matters.

• Pin the version. Always use an exact pinned version. Never use a floating tag such as ecc-agentshield@latest, which can execute changed third-party code at any later run.
• Keep it project-local. Install into the project with a lockfile rather than globally, so the exact audited version is pinned in the repository and the install is reviewable in version control.
• Use an isolated work tree. Run the scan against a version-controlled or throwaway working tree so any file changes are reviewable and reversible.
• Hold off on --fix. Read the proposed changes before letting the tool apply them. --fix modifies your configuration — that authority should be deliberate.
• Hold off on init. init writes a starter configuration. Read what it would create before letting it write to your config directory.
• Scope the API key. If you enable deep analysis, use a limited API key rather than your primary credential.

Install (project-local, pinned)

bash

# Pin an exact version into the project. The lockfile records the exact tree.
npm install --save-dev ecc-agentshield@1.2.3

Do not use a global install (npm install -g) and do not use an unpinned npx ecc-agentshield@latest. Both defeat the pinning that makes this scan reproducible and reviewable.

Run (pinned, local)

bash

# Scan the current project's .claude directory using the pinned local install
npx --no-install ecc-agentshield scan --path .claude

npx --no-install refuses to fetch anything that is not already installed, so the scan cannot silently pull a different version at run time.

When to Activate

• Setting up a new Claude Code project
• After modifying settings.json, CLAUDE.md, or MCP configs
• Before committing configuration changes
• Onboarding to repo with existing configs
• Periodic security hygiene checks

What It Scans

• CLAUDE.md — Hardcoded secrets, auto-run instructions, injection patterns
• settings.json — Overly permissive allow lists, missing deny lists
• mcp.json — Risky MCP servers, hardcoded env secrets
• hooks/ — Command injection via interpolation, data exfiltration
• agents/ — Unrestricted tool access, missing model specs

Command reference, severity grades, and critical/high finding detail live in references/commands-and-findings.md.