Langflow RCE (CVE-2026-33017)

Overview

This workflow finds Langflow and sorts each host by whether its version falls in the CVE-2026-33017 range. Langflow is the visual builder teams stand up for LLM graphs. CVE-2026-33017 is code injection that CISA listed on KEV. A below-1.9.0 API still answers /api/v1/version. Give it hostnames, IP addresses, or URLs you are authorised to test. The workflow fingerprints the product from the public page (title match on Langflow), reads the version from /api/v1/version with a fallback to the root body, and compares that number to the range table in the graph. Every host comes back affected or not, so a fleet advisory becomes an evidenced list for exposure management. CVE-2026-33017 is on CISA KEV (known exploited). Version flag is first. Hosts in range get a second node that only asks the unauthenticated build path to sleep. It does not send an RCE payload.

Run it on a schedule when the estate changes. New hosts and delayed patch windows are the reason the same check is worth more the second week than the first.

Pipeline

1. Read the target list. Hosts, URLs, or ranges, one per line, become the scope.
2. Fingerprint each host as Langflow from the public page title, then GET /api/v1/version and parse the version. A host counts as affected only when the fingerprint matches and the version is inside the encoded range.
3. Confirm only those flagged hosts with a time-based sleep on the unauthenticated build path.
4. Collect the per-host rows: product match, version, affected flag, confirmed flag, and any fetch error.
5. Emit the summary counts: hosts checked, product hits, affected, confirmed, and errors.

Inputs

• Target list. Hostnames, IP addresses, CIDR ranges, or URLs, one per line. JSONL with a url field also works. Point it at scope you are authorised to test. Fingerprint is read-only GET. Confirm is a sleep-only probe on the unauthenticated build path, and only after the version flag.

Outputs

• results.jsonl. One row per host: URL, whether Langflow was detected, the version read, the affected flag, and whether the sleep confirm succeeded.
• summary.json. Counts across the list: targets, product hits, affected, and errors.

Integrations

• HTTP. Unauthenticated GET to the product root and /api/v1/version. Confirm is a sleep-only probe on the unauthenticated build path, only on flagged hosts.

Sample output

The records below are illustrative and do not come from a real run. They show one Langflow host inside the CVE-2026-33017 range, one patched host, and one host that is not the product.

{"cve": "CVE-2026-33017", "total_targets": 3, "affected": 1, "not_affected": 1, "errors": 1}

FAQ

Which Langflow versions does this flag?

every build before 1.9.0. 1.9.0 and later is the floor the graph treats as not affected. The compare uses the version the product itself served.

Does a flagged host mean CVE-2026-33017 was exploited?

No. A flagged row is patch-lag against the range table. Confirm only asks the unauthenticated build path to sleep. Treat a confirmed row as evidence the injection still answers, not as proof someone already used it.

Is this check safe on production?

Yes. Fingerprint is read-only GET. Confirm is a sleep-only probe on the unauthenticated build path, only on hosts already in range. It does not send an RCE payload.

Does the check need credentials?

No. It fingerprints the public page and reads /api/v1/version the way an external scanner would.

What is CVE-2026-33017?

Code injection in Langflow. CISA listed it on KEV. The check flags every build before 1.9.0. This workflow flags by version, then confirms with a sleep-only probe.

Related workflows

• LiteLLM SQLi (CVE-2026-42208). Reach for it when the next LLM-proxy flag is SQL injection.
• Metabase SQLi (CVE-2026-72898). Reach for it when the next internal app is a BI console.
• Web AI Agent Audit. Reach for it when you want probe coverage of an agent page, not a Langflow version flag.