Please do not report security vulnerabilities through public GitHub issues.

Report them to security@sanity.io. Include a description of the issue, steps to reproduce, and the potential impact. You should receive a response within a few business days.

This harness executes code generated by AI agents: it writes agent-emitted files to a sandbox project directory, runs npm install on agent-chosen dependencies, and serves the result in a local dev server. Run it in an environment you trust with that. Reports about the harness writing or reading outside its sandbox directories are in scope and appreciated.

A test's config.js can declare an mcp block whose command and args are spawned as a child process during a run (see "Per-test MCP" in the README). Those values are executed verbatim. Treat every test config — your own and any you copy from elsewhere — as trusted code, the same as any other file you'd node. Do not run tests whose config.js you have not reviewed.