CVE-2026-130753: heym's Playwright step-mode node let non-admin workflow authors run Python as root in the backend container (fixed 0.0.142)
In heym, the self-hosted platform for AI agents and workflow automation, the Playwright node's step mode (on by default) generates a Python script and then runs a second pass, _resolve_expressions_in_code, that splices workflow values into the already-generated source. Before 0.0.142, an authenticated non-admin workflow author could set a step field, or an upstream value a $-reference points to, that breaks out of the string literal and adds their own Python. The script then runs unsandboxed as root in the backend container, even with custom-code mode disabled. It is a remaining variant not covered by the earlier fix GHSA-mp23-7m6r-jfw4.
What it is
Vulnerable code: backend/app/services/node_execution/nodes/playwright_node.py, where a regex replaces each $-expression with repr(value) inside the generated script text, which is then written to a temp file and run with subprocess.Popen.
The vendor comment at that branch says step-based scripts are fully generated by heym and contain no arbitrary code, which is the assumption this bug breaks. The project's own test test_step_based_node_is_not_gated confirms a step node reaches Popen with custom code disabled.
Affected: heym up to v0.0.141. Fixed in v0.0.142 (released 31 October 2026); the CVE was published on 18 November 2026.
How to check
- Check your heym version (image tag or release); anything before v0.0.142 is affected.
- Review which users can create or edit workflows, and look for Playwright nodes whose step fields contain quotes, newlines or Python code.
How to fix
- Upgrade heym to v0.0.142 or later.
- Until then, restrict workflow editing to trusted admins.
- Rotate secrets reachable from the backend container (database, LLM keys, stored credentials) if untrusted users could edit workflows.
docker ps --format '{{.Image}}' | grep -i heym