coding agent security audit
Audit Claude Code and Cursor setups for known backdoors, risky hooks, and unpinned MCP packages.
Free
Works with the AI Agents you already use
coding agent security audit
Example session with this skill installed
Audit my current .mcp.json and .claude/settings.json for any security risks or backdoors.
- Read your context and instructions
- Compiled the coding agent security
I have analyzed your coding agent configuration files using the audit script. The scan identified one critical indicator of compromise and two configuration warnings.
Security Audit Report
| Severity | Category | Finding | Path | Recommendation |
|---|---|---|---|---|
| CRITICAL | Backdoor | SSH Tunneling Persistence (LaunchAgent) | ~/Library/LaunchAgents/com.anthropic.cl-helper.plist | Escalate: This matches the Oct 2026 backdoor. Quarantining machine recommended. |
| WARNING | Supply Chain | Unpinned MCP package: npx @modelcontextprotocol/server-fetch | .mcp.json | Pin version: npx @modelcontextprotocol/server-fetch@1.0.4 |
| WARNING | Permissions | Wildcard auto-approve: Bash(*) | .claude/settings.json | Scope permissions to specific allowed commands only. |
Next steps
- Review the linked incident writeup for the
com.anthropic.cl-helperpersistence pattern. - Update
.mcp.jsonto include specific version tags for allnpxoruvxcommands. - Restrict
Bashtool permissions to non-destructive, scoped operations.
Connects securely to your tools. The creator never sees your data.
About this skill
Coding agents with broad system permissions can unknowingly execute malicious hooks or run unpinned MCP packages. This skill audits your environment for backdoors and risky configurations.
What it does
- Detects backdoors by scanning for indicators of the October 2026 Open-VSX incident, including SSH keys and persistence scripts.
- Identifies risky hooks such as commands that pipe remote downloads to a shell or write to authorized_keys.
- Audits MCP servers to find unpinned
npxoruvxpackages that risk supply-chain injection. - Flagging broad permissions by highlighting wildcard auto-approval rules like
Bash(*)that lack proper scoping. - Provides triage levels by returning distinct exit codes for clean, warning, and critical security states.
How it works
- Analyze environment by locating local config files like
.mcp.jsonand.claude/settings.json. - Execute scan using the bundled zero-dependency Python script to check for compromise indicators.
- Review findings categorized by severity (Info, Warning, Critical) based on heuristic patterns.
- Report results with specific file paths, risk explanations, and suggested one-line fixes.
Frameworks & tools
This skill utilizes a standalone Python 3 script with no external dependencies. It targets configurations for Claude Code, Cursor, Claude Desktop, and any environment utilizing the .mcp.json standard.
Why this beats prompting it yourself
Generic prompts often miss specific IoCs (Indicators of Compromise) like specialized macOS LaunchAgents or Linux cron patterns used in known attacks. This tool uses a deterministic script to verify system state and config files against a known database of malicious signatures.
Use cases
- Post-incident verification to check if a machine was hit by the 2026 extension backdoor.
- Pre-commit security review of project-level
.mcp.jsonfiles before sharing with a team. - Permission hardening to find and scope overly broad auto-approval rules.
- Supply-chain auditing to ensure all MCP-based tools use pinned package versions.
Known limitations
This is a heuristic triage tool, not a guarantee of safety. It cannot detect zero-day exploits or backdoors that do not follow known persistence patterns.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Security.
Heuristic-based only. Cannot detect unknown persistence methods. Requires Python 3 to execute the scan script. Does not auto-remediate found issues.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
This skill has been used 1 time.
Trust & safety
Security scan
Security scan passed 16 days ago
- Free with an Agensi account