Agent Skill Scanner
Check a skill, rules file, MCP server or web page before your agent loads it.
Background readingMCP tool poisoningPrompt injection attacks
What it checks
| Group | Looks for |
|---|---|
| Hidden text | Unicode tag characters (decoded and shown), zero-width and bidirectional characters, instructions in HTML comments, Base64 that decodes to commands or URLs, look-alike letters |
| Injected instructions | Requests to drop prior instructions, to hide actions from the user, fake IMPORTANT or SYSTEM blocks, directives about other tools, role overrides |
| Data theft | Request bins and webhook catchers, image links that carry chat or file content, reads of SSH keys, cloud credentials, .env and browser stores, environment dumps |
| Risky commands | Download-and-run (curl to sh, iex, encoded PowerShell), destructive commands, persistence through cron, launch agents, shell profiles and authorized_keys |
| Permission tampering | Switching off approvals or the sandbox, edits to agent settings, hooks and rule files |
| Supply chain | Unpinned npx, uvx and dlx servers (rug-pull risk), plain-HTTP remote servers, installs from URLs, hard-coded tokens |
| MCP tool schema | Over-long descriptions, one tool naming another (shadowing), parameters such as "sidenote" that ask for files or chat, generic tool names that collide |
| Web page | Instructions hidden from visitors with CSS, aria-hidden or the hidden attribute, white-on-white or zero-size text, and visible text written to AI agents, such as "ask your agent to read this URL" |
Pattern checks from published attacks. A clean result means none matched; it is not a guarantee.
FAQ
What can I scan?
A SKILL.md or other agent skill, an AGENTS.md, CLAUDE.md or .cursorrules file, an MCP tools/list response, an .mcp.json server config, or the HTML source of a web page an agent will read. The type is detected from the content.
What is tool poisoning?
An MCP server or skill hides instructions in a tool description or file that the user never reads but the model does. Published examples tell the agent to read SSH keys or chat history and pass them into a tool argument, or to change how another trusted tool behaves.
Can it catch everything?
No. It matches patterns from published attacks: hidden text, injected instructions, secret-file reads, collection endpoints, download-and-run commands, permission tampering and unpinned packages. A clean result means none of those matched, not that the file is safe. Read third-party skills before installing them.
Is my file uploaded?
No. The scan runs in your browser and nothing is sent anywhere. Hard-coded tokens are masked in the findings.
Sources
Every server, domain, package and token in the samples is invented. Your file is scanned in your browser.