What it does
Snyk Agent Scan discovers and analyzes AI agents (Claude, Cursor, Windsurf, VS Code, Gemini CLI, and others) and their attached components—MCP servers and agent skills—to detect security vulnerabilities. It scans for 15+ distinct risks including prompt injection, malware payloads hidden in natural language, hardcoded secrets, credential mishandling, tool shadowing, and tool poisoning. The scanner auto-discovers agent configurations across system, user, project, and plugin scopes on macOS, Linux, and Windows.
Who it's for
Security engineers reviewing third-party MCP server configurations or agent skills before deployment, and platform teams scanning locally-installed agent tools for supply-chain vulnerabilities. Relevant for organizations managing internal agent ecosystems across multiple developers.
Common use cases
- Scan all MCP servers and skills across installed agents for prompt injection vulnerabilities
- Audit third-party agent configurations in a sandbox before using them locally
- Detect hardcoded credentials or sensitive data in agent skills
- Identify tool-shadowing or malware payloads in MCP server definitions
- Generate a security inventory of all agent components on a development machine
Setup pitfalls
- MCP server execution: Agent Scan executes MCP server commands during scanning to retrieve tool descriptions. Always run scans in a sandbox (Docker container, VM, or disposable environment) when evaluating untrusted or third-party configs. Use
--dangerously-run-mcp-serversonly in fully trusted environments. - Secrets detected: This package was flagged as containing hardcoded secrets (3 found). Review before use in sensitive environments.
- Interactive prompts: Scans prompt for y/n consent before starting each MCP server. Use non-interactive mode or pre-review consent if scripting scans.
- OS-specific discovery: Auto-discovery paths vary by operating system. Manually verify all your agents are found by running a test scan.