Skip to main content

CVE-2026-33017: Langflow's Public Flow Endpoint Runs Unauthenticated Code, and No Stable Release Fixes It

· 7 min read
Ankush K
Co-founder and CTO, Dwaar
CVE-2026-33017: Langflow's Public Flow Endpoint Runs Unauthenticated Code, and No Stable Release Fixes It

If your Langflow instance is reachable from the internet, the question today is not when you will upgrade. No stable langflow release on PyPI carries the fix, and the advisory describing an unauthenticated route to exec() has been public for three days. That instance probably holds your API keys, so treat this as containment.

ItemDetail
IdentifierCVE-2026-33017, GHSA-vwmf-pq79-vjvx
Affectedpip package langflow <= 1.8.1, per the March 17 Advisory Database entry. That includes the newest PyPI release as of March 19, 1.8.1
Fixed inNo patched version listed in the Advisory Database as of March 19, and no stable langflow release carries the fix. It exists as commit 73b6612 and in the langflow-nightly dev builds
SeverityCritical, CVSS 4.0 base 9.3, in the GitHub advisory
CISA KEVNot listed in the March 19 catalog
DisclosedRepository advisory March 16, 2026. GitHub Advisory Database March 17, 2026

The public check approves one flow and the server builds another​

Langflow lets anonymous visitors run public flows, so POST /api/v1/build_public_tmp/{flow_id}/flow has no Depends(get_current_active_user), unlike the authenticated build endpoint in the same file. That part is intended. The endpoint uses a client_id cookie to derive a deterministic flow ID and verify_public_flow_and_get_user to confirm the flow is public.

The bug is one optional body parameter. Per the advisory, the endpoint also accepts data and forwards it through start_flow_build(data=data), so the graph gets built from the caller's JSON instead of the stored flow.

Flow nodes carry Python. The node's code reaches exec() in prepare_global_scope, in src/lfx/src/lfx/custom/validate.py, with no sandbox. Top-level assignments execute during graph building, so the attacker's code runs before the flow itself does, with the privileges of the server process.

The preconditions are weaker than they look. The advisory asks for one public flow, its UUID, and a client_id cookie holding any string. With AUTO_LOGIN=true, which it gives as the default, an attacker can get a superuser token from /api/v1/auto_login and create a public flow. "We have no public flows" is no defense on a default install.

The fix commit, authored March 13, deletes the parameter and always passes data=None, so the definition loads from the database. The advisory separates this from CVE-2025-3248, which was fixed by adding authentication to /api/v1/validate/code. That remedy cannot work here, because this endpoint is meant to be open.

The advisory leaves little to work out​

The repository advisory page shows publication on March 16 at 12:20 UTC, and the Advisory Database entry is stamped March 17 at 20:05 UTC. Depending on which clock you use, the full detail has been public for two to three days.

Nobody needs to reverse a patch here. The advisory names the endpoint, the parameter and the sink, and it reports its own proof of concept, confirmed in six of six runs against 1.7.3.

The March 19 KEV catalog does not list this CVE. That tells you what CISA has added so far, and nothing about the risk to your instance. Absence from the catalog does not mean no exploitation.

What to do today​

  1. Find every instance. Any langflow at or below 1.8.1 is in the advisory's range, which includes the newest PyPI release.
  2. Take it off the internet. Put it behind a firewall or an authenticating reverse proxy. If it must stay reachable, block /api/v1/build_public_tmp. You lose anonymous public flows, and that is the right trade this week. The advisory lists no workarounds, so this is my own judgement.
  3. Check the code, not the version string. If you build from source, confirm that commit 73b6612 (#12160) is in your tree and that data is gone from build_public_tmp. The nightly builds carry the commit, but I would not treat a nightly as your production upgrade path.
  4. Check AUTO_LOGIN. This is my inference from the advisory's prerequisites, not a vendor workaround. I would not leave it on for anything an untrusted network can reach.
  5. Rotate secrets on anything that was exposed, and treat exposure since March 18 as possible compromise. API keys, database passwords, cloud credentials.
  6. Hunt, using what the advisory describes. These are my reading of its attack chain and impact list. Do not assume you are early: public honeypot reporting from researchers, dated March 19, describes exploitation attempts starting March 18, about 20 hours after the Advisory Database entry, including environment variable exfiltration.
    • POST requests to that path whose body carries a data object with node code. If you do not log request bodies, access logs cannot rule this out.
    • Requests to /api/v1/auto_login followed by POST /api/v1/flows/ from addresses you do not recognize, which is the advisory's route to creating a public flow.
    • Child processes of Langflow you cannot explain. The advisory's proof of concept shells out to id.
    • Outbound connections from the host you cannot explain. The advisory's impact list includes reverse shells and environment variable exfiltration.

The patch window is the wrong control​

A patch deadline assumes a fixed release exists. As of March 19 that is not true for Langflow's stable channel. The advisory is public, and no stable release carries a commit authored six days ago.

My position: for internet-reachable AI tooling, plan on a patch window measured in hours, so your real controls are the ones already in place when the advisory lands. Agent builders stay private by default, and someone has the authority to pull one off the network the same day.

Two things make this class of tool worse. First, connecting things is its job, so I expect its environment to collect credentials, model provider keys among them. The advisory's impact list names environment variables holding "API keys, database credentials, cloud tokens". Give each deployment scoped keys and practise rotating them.

Second, a flow definition is code. CVE-2025-3248 sits in the same family of caller-supplied code reaching exec(), and CISA's KEV catalog has listed it since May 5, 2025. Any endpoint that accepts a flow from a caller is a remote execution API, whatever its name says.

Do not wait for KEV either. A listing is a strong signal, and absence tells you nothing. CVSS Is Not Risk covers how to weigh those signals.

Pick one agent builder you run today. Write down who can reach it and every credential in its environment, then time how long rotating all of them takes. If that takes longer than the three days since this advisory went up, fix that first.

References​