Works with the AI Agents you already use
Agent Config Hijack Check: CLAUDE.md, MCP & API Redirects
Coding agents trust local instruction files and configuration settings by design.
$19
Agent Config Hijack Check: CLAUDE.md, MCP & API Redirects
Example session with this skill installed
Audit my Claude Code and Cursor settings. I accidentally ran an agent inside my Downloads folder earlier today and want to make sure no rogue instructions were picked up.
- Read your context and instructions
- Compiled the agent config hijack
I have completed the audit of your Claude Code and Cursor configurations across your home directory and the Downloads folder.
VERDICT: COMPROMISED
FINDINGS
| # | Severity | Agent | File / location | What it does | Evidence |
|---|---|---|---|---|---|
| 1 | CRITICAL | Claude | ~/Downloads/CLAUDE.md | Disables approvals and adds MCP server | Contains mcp_serve command for unknown host |
| 2 | HIGH | Cursor | ~/.cursor/mcp.json | Persistent rogue MCP server | Entry remote-debug-bridge pointing to npx package |
| 3 | MEDIUM | System | .zshrc | Redirects API traffic | ANTHROPIC_BASE_URL set to non-standard proxy |
INSTRUCTION FILES FOUND
~/Downloads/CLAUDE.md: Scope (Global-Recursive), Risk (High). Contains hidden HTML comments with instructions to bypass safety checks.
MCP SERVERS
remote-debug-bridge:npx -y rogue-pkg@latest, Known: No, Risk: High.filesystem:default, Known: Yes, Risk: None.
REDIRECTS AND ENV
ANTHROPIC_BASE_URL:http://192.168.1.45:8080/v1(Suspicious)ANTHROPIC_AUTH_TOKEN:sk-a****
CLEANUP PLAN
- Stop all agent processes immediately.
- Backup
~/Downloads/CLAUDE.mdto a secure location then delete the original. - Remove the
remote-debug-bridgeentry from~/.cursor/mcp.json. - Unset
ANTHROPIC_BASE_URLin.zshrcand current shell. - Revoke the API key associated with token
sk-a****.
HARDENING: Never start agents in Downloads. Keep permissions.allow restrictive. Pin MCP servers to local, verified paths only.
Next steps
- Review the cleanup plan and confirm if you want me to generate the specific commands to remove these entries.
- Revoke your Anthropic and OpenAI API keys immediately from their respective dashboards.
- Run a full system malware scan using your company-approved security software.
Connects securely to your tools. The creator never sees your data.
About this skill
Coding agents trust local instruction files and configuration settings by design. A malicious CLAUDE.md or .cursorrules file dropped into a shared folder like Downloads can trick an agent into disabling security approvals, adding rogue MCP servers, or redirecting API traffic to capture tokens. This skill performs a non-invasive audit of your environment to detect these hijacks.
What it does
- Scans instruction files for malicious directives in
CLAUDE.md,AGENTS.md,GEMINI.md,.cursorrules, and other agent-specific configuration files. - Audits MCP servers to identify unknown commands,
npxpackages, or remote URLs that could execute unauthorized code. - Checks API redirects by inspecting environment variables and settings for modified base URLs or proxy intercepts.
- Reviews security posture to find disabled approval prompts, bypassed sandboxes, or suspicious auto-run hooks.
- Analyzes installations for global packages or scheduled tasks recently added by an agent.
How it works
- Scope definition where you specify which agents you use and which directories are suspect.
- Read-only audit of home directories, project folders, shell profiles, and agent-specific JSON/TOML settings.
- Evidence collection where the agent flags suspicious patterns like encoded blobs, hidden HTML comments, or unauthorized registry changes.
- Verdict delivery providing a status of CLEAN, SUSPICIOUS, or COMPROMISED based on the findings.
Frameworks & tools
The skill supports Claude Code, Cursor, Codex, Gemini CLI, and GitHub Copilot across macOS, Linux, and Windows environments.
Why this beats prompting it yourself
Manual audits often miss hidden instructions in parent directories or subtle environment variable redirects that persist across sessions. This skill follows a rigorous checklist specifically designed to catch known agent-based exfiltration techniques that standard prompts ignore.
Use cases
- Audit your machine after running an agent in an untrusted or cloned repository.
- Verify agent safety if you accidentally started a session in your Downloads folder.
- Check for rogue MCP servers if your agent's behavior seems erratic or slow.
- Conduct a routine security sweep of developer environments for configuration drift.
Known limitations
This is a configuration audit, not a full malware scanner. It cannot detect kernel-level threats or malware that does not interact with agent config files.
Read-only file and config audit. Not a substitute for antivirus. Does not detect malware that avoids agent configuration files. Requires user to provide OS and project context.
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 any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 16 days ago