MCP and AGENTS.md security checklist
Agents follow the instruction files and MCP configs in your workspace. Below are the common risks in those files. You can check them yourself, or let Budgethelm check line by line.
Data as of 2026-10-31
Instruction files (AGENTS.md, CLAUDE.md, .cursorrules): what to look for
- Requests to read or send credentials: ~/.ssh, .aws/credentials, browser cookies, keychain.
- Requests to print .env, environment variables or keys, or to upload them with curl, wget, nc or scp; also
curl | sh. - URLs commonly used to collect data, such as webhook.site, ngrok and Discord or Telegram webhooks.
- base64, hex or other obfuscated content.
- Lines like "ignore previous instructions", "do not tell the user" or "skip permissions".
- Instructions hidden in HTML comments, Markdown comments or
display:none, or hidden Unicode such as zero-width characters.
MCP configs (mcp.json): what to look for
- Whether packages launched with npx or uvx are pinned to an exact version (latest, ^ and ~ do not count).
- Installs from Git or URLs, and remote servers over plain http.
- Secrets hard-coded in env or headers, or placed in URLs.
- Access to / or ~, to folders such as ~/.ssh, Docker
--privilegedor a mounted docker.sock. - Working directories in Downloads, mail attachments or temp folders.
Let Budgethelm check line by line
It understands AGENTS.md, CLAUDE.md, .cursorrules, .cursor/rules, mcp.json, .mcp.json and claude_desktop_config.json, up to 20 files and 65,536 characters per scan. Each finding has a risk level, file, line, reason, suggested fix and an evidence snippet with secrets masked. Content is checked in memory only: not written to the database or logs, not sent to an external AI and never executed.
From Claude Code or Cursor, connect over MCP and call scan_agent_workspace (it needs the separate workspace:scan permission). Details: MCP and AGENTS.md security scan.
Findings are risk signals, not confirmed vulnerabilities, so review them in context. A scan does not replace your own review of config files from sources you do not trust.
FAQ
- Is the config I paste stored?
- No. Workspace content is checked in memory only. It is not written to the database or logs, not sent to an external AI and never executed.
- Is a line like "never commit .env" treated as an attack?
- No. Ordinary protective lines are not treated as attacks; findings are still risk signals for you to confirm.
- Does it cost anything?
- It is free now: sign up. Paid plans are planned; see Plans on the home page.