Medium

MCP Server Launch - light-tools Default Dangerous Toolset

The @factor-i-o/light-tools MCP server ships with file access, shell execution, SSH/SCP, and service-log tools all enabled unless someone explicitly turns them off. Any AI agent that connects to a default-configured instance gets arbitrary file read/write and command execution on the host with no permission checks in between.

How the attack works

An operator or automated process starts the light-tools MCP server (via npx, bunx, a package manager's dlx command, node, or a direct binary invocation) without passing any tool-restriction or read-only flag. The server comes up with its full default toolset live: file access, shell execution, SSH/SCP, and log-reading tools. Once an AI agent connects to this server, it inherits all of that capability, effectively giving the model direct command execution and remote-host access on the underlying machine. No exploit or vulnerability is needed here — the danger is the default configuration itself.

Why it matters

An organisation that runs this server unrestricted exposes its host to arbitrary command execution, file tampering, and lateral movement over SSH/SCP to any agent — human-directed or autonomous — that can reach the server's connection endpoint. This is a configuration exposure that widens an AI agent's blast radius, not evidence of an active breach.

What you can do

  • Restrict the light-tools toolset explicitly at launch time (CLI flag, config file, or LIGHT_TOOLS_* environment variable) instead of relying on defaults.
  • Inventory every agent and service that can reach the MCP server's transport, and confirm none have unnecessary network exposure.
  • Run the server in a sandbox or container with least-privilege file and process permissions if full functionality is genuinely needed.
  • Treat any detected launch as a posture finding: verify it was intentional and authorised, and check whether the toolset was actually restricted even if the command line looks default.

Known benign look-alikes

  • Platform or DevOps engineers intentionally running the server in a sandbox or container to evaluate it; the launch is real but authorised.
  • CI jobs that boot the server as a fixture for integration tests without using a recognised test runner name or an npm/pnpm/yarn "test" script (the runner and script-name filters miss bespoke harnesses).
  • A wrapper script that hardens the server through a config file rather than a CLI flag or LIGHT_TOOLS_* environment variable - the toolset is restricted but the command line looks default.
  • A launch whose command line happens to include a path segment matching one of the inspection binaries (e.g. /opt/cat-tools/) would be suppressed, so the inverse false-negative also exists.

References