Grok Workhorse runs untrusted, model-driven code on your machine. Sandbox escapes and credential leaks are the bugs we care about most.

Please do not open a public issue for security problems. Use GitHub's private vulnerability reporting on this repository instead (Security tab → "Report a vulnerability").

Please include:

• the version or commit, OS/kernel, backend and backend version
• what a worker (or a malicious repo or prompt) can do that it should not be able to do
• a minimal reproduction, ideally a mock scenario (test/mock/scenarios/*.json) or the tool calls involved

You can expect an acknowledgement within 7 days. Once a fix is available we will publish an advisory and credit you unless you prefer otherwise.

In scope:

• escaping the outer sandbox or reading hidden paths (secret store, daemon token, other tasks' data, the main clone's working tree)
• getting a provider key or other secret into worker-visible output, logs or results
• bypassing the repo, test-command or clone-host allowlists through the MCP interface
• making the daemon report a false verdict (for example hiding commits)
• privilege escalation to root or to other users

Known limitations that are documented in docs/security.md (for example network access for model-run commands on backends without an inner sandbox) are not vulnerabilities by themselves. Ways to widen their impact are welcome as reports.

Only the latest release receives security fixes during the 0.x series.