← All advisories
CVE-2026-134360highCode execution2026-11-21

Heym: any workflow user could read backend files (incl. the Kubernetes service-account token) via Playwright file:// steps, collaborators could steal the owner's credentials via error_workflow_id, and the Python tool sandbox could be escaped (CVE-2026-134360, CVE-2026-110873, CVE-2026-55376)

Three more flaws in Heym, the self-hosted agentic workflow platform. Newest (CVE-2026-134360, fixed 0.0.148): the Playwright node's navigate step passed its URL to page.goto() without checking the scheme, and step-based scripts run inside the backend container instead of the sandbox, so any signed-in user who can run a workflow could navigate to file:///… and use getText to read known-path files such as the Kubernetes ServiceAccount token or Docker secrets; AI steps passed model-chosen URLs the same way. CVE-2026-110873 (0.0.55 to 0.0.120, fixed 0.0.121): a shared collaborator could point a workflow's error_workflow_id at their own workflow, and when the parent errored the owner's decrypted credentials were injected into it. CVE-2026-55376 (fixed 0.0.40): the blacklist-based Python tool sandbox could be escaped through allowed standard-library modules, giving command execution as the backend user.

01.

What it is

CVE-2026-134360 (GHSA-hmfh-wpg2-vff4): services/playwright_code_generator.py passes navigate URLs straight to page.goto(); nodes/playwright_node.py runs step-based scripts in the backend container. The reporter confirmed reading a mounted ServiceAccount token as a newly registered non-admin user on the 0.0.145 release image; 0.0.146 and 0.0.147 are also affected. In the release image the browser runs as root. With the sandbox enabled, that token's Role can create/delete pods in Heym's namespace, which the advisory chains to reading secrets such as ENCRYPTION_KEY and DATABASE_URL. Reported by amasolov.

CVE-2026-110873 (GHSA-c3f4-2mrj-jfqr): error_workflow_id was missing from OWNER_ONLY_EXECUTION_FIELDS in backend/app/api/workflows.py, so PUT /api/workflows/{id} let a collaborator change it. The error runner falls back to the owner's credential context, so an HTTP node in the attacker's error workflow could exfiltrate every credential the owner holds.

CVE-2026-55376 (GHSA-wcgw-9hfw-f6f2): backend/app/services/python_tool_executor.py allowed standard-library modules whose public attributes expose OS, process and filesystem access, then ran the user's tool with exec(). Needs a user who can create or edit a workflow with an Agent custom Python tool.

NVD published CVE-2026-134360 on 21 November 2026; CVE-2026-110873 and CVE-2026-55376 were published by NVD on 14 and 13 October 2026.

02.

How to check

  1. Check your Heym version (container image tag ghcr.io/heymrun/heym:<version>).
  2. Review workflows for Playwright navigate steps with file:// URLs and for error_workflow_id values that point at workflows owned by someone else.
  3. On Kubernetes, check what the backend pod's ServiceAccount is allowed to do.
03.

How to fix

  1. Upgrade Heym to 0.0.148 or later (also covers the 0.0.121 and 0.0.40 fixes).
  2. If untrusted or self-registered users could run workflows on an older version, rotate the Kubernetes ServiceAccount token, ENCRYPTION_KEY/SECRET_KEY, database credentials and the credentials stored in Heym.
  3. Disable open self-registration on shared instances and only share workflows with trusted collaborators.
docker ps --format '{{.Image}}' | grep -i heym
# Kubernetes
kubectl get pods -A -o jsonpath='{range .items[*]}{.spec.containers[*].image}{"\n"}{end}' | grep -i heym
04.

Sources

Related on VibeVulns
The weekly email

One email a week. Only what changed.

Every new security hole in AI coding tools, MCP servers, app builders and the stacks they generate (Next.js, Supabase, Node, PHP), each with how to check your app and the fix. No roundups of things you already read.

Free, one email a week. Read the latest issue first if you like. One-click unsubscribe in every email. We never sell or share addresses.