# CVE-2026-9198: IBM Langflow OSS, is it exploitable where you run it

Ledger entry, disclosed 2 Jul 2026, sources read 19 Nov 2026

**The short answer.** CVE-2026-9198 affects IBM Langflow OSS and was disclosed 2 Jul 2026 with a CVSS score of 9.8. Exposure needs these conditions: Langflow OSS runs a version from 1.0.0 through 1.10.0; auto-login is enabled (LANGFLOW_AUTO_LOGIN set to true), which IBM calls the default in the affected versions. CISA added it to the Known Exploited Vulnerabilities catalog on 4 Aug 2026.

## What the flaw is

Langflow is a tool for building AI agent and workflow applications. CVE-2026-9198 chains two flaws. First, an unauthenticated auto-login endpoint (/api/v1/auto_login) issues superuser bearer tokens to any caller on the network. Second, a code validation endpoint (/api/v1/validate/code) runs submitted Python when it defines a function: decorators, default arguments and annotations are executed at definition time.

An attacker who can reach the server gets a superuser token from the first endpoint and then uses it to run code through the second. The result is remote code execution as the Langflow process, with no login and no user interaction. IBM says this affected default deployments where auto-login was enabled and the validation endpoint was reachable. IBM rates it 9.8 (CVSS 3.1 base score; network, low complexity, no privileges). The weakness is CWE-94, code injection.

## Are you exposed: each condition and how to check it

| Condition for exposure | How to check it on your system |
| --- | --- |
| Langflow OSS runs a version from 1.0.0 through 1.10.0 | On the host run pip show langflow, or pip list \| grep -i langflow. In a container run docker exec CONTAINER pip show langflow. 1.10.1 is the first fixed release. |
| auto-login is enabled (LANGFLOW_AUTO_LOGIN set to true), which IBM calls the default in the affected versions | Check the effective value: env \| grep LANGFLOW_AUTO_LOGIN, the .env file passed with --env-file, and for containers docker inspect --format '{{.Config.Env}}' CONTAINER. Langflow's current documentation says newer official packages and images ship with it set to false, and that an explicit true in the environment or .env file overrides that default. Treat an unset value on 1.10.0 or earlier as enabled. |
| the server is reachable over the network by the attacker, so both API routes can be called | Run ss -ltnp \| grep -i langflow to see the bind address and port (Langflow documentation uses 7860 in its examples). Then check firewall or security group rules and any reverse proxy for that port from an untrusted network. Do not call the auto_login route on a live system to test it: on an affected version it issues a superuser token. |
| no authenticating reverse proxy sits in front of the server | Review the reverse proxy configuration. Langflow's documentation says to deploy behind a reverse proxy with authentication and never to expose its ports directly to the internet. |

## Affected and fixed versions

| Product line | Affected | First fixed | Notes |
| --- | --- | --- | --- |
| Langflow OSS (IBM bulletin; PyPI package langflow) | 1.0.0 through 1.10.0 | 1.10.1 (released 23 Jun 2026) | IBM bulletin published 2 Jul 2026; NVD configuration reads 1.0.0 up to but excluding 1.10.1 |

The 1.10.1 release notes list two matching fixes without citing the CVE: 'do not execute code in validate_code' and 'stop issuing 365-day superuser token via auto_login'.

## If the fix cannot go in yet

- Upgrade Langflow OSS to 1.10.1 or later. IBM strongly recommends this and lists no workarounds or mitigations. Cost: a change to authentication behavior, so read the next item before restarting an existing install.
- Exposure-reducing step: set LANGFLOW_AUTO_LOGIN=false and set LANGFLOW_SUPERUSER_PASSWORD, as Langflow's documentation recommends. IBM limits the flaw to deployments with auto-login enabled, but does not list this as a workaround, so upgrade as well. Cost: users must sign in to the editor. On an existing auto-login install, prepare the account first (Langflow documentation): back up the database and configuration directory; keep the same database, configuration directory, encryption key and superuser username; set a known password on the existing account with PATCH /api/v1/users/{user_id} (the self-service reset endpoint needs the old password, so it does not suit an unknown generated one) and verify it with POST /api/v1/login; then set LANGFLOW_AUTO_LOGIN=false, LANGFLOW_SUPERUSER to that username and LANGFLOW_SUPERUSER_PASSWORD, and restart. Creating a different superuser does not transfer the flows.
- Disabling auto-login does not revoke sessions or API keys issued earlier (Langflow documentation), and the 1.10.1 release notes name a 365-day superuser token. Plan to revoke or rotate credentials on any instance that was exposed.
- Do not expose Langflow to untrusted networks and put an authenticating reverse proxy in front of it. Cost: clients must authenticate to the proxy.
- CISA's KEV entry lists the required action: apply vendor mitigations, follow the forensic triage requirements it cites, and discontinue use of the product if mitigations are unavailable.

## What to look for

- Search the reverse proxy or web server access log, and Langflow's own uvicorn access log, for the two routes. Langflow publishes no log lines for this CVE; these two are illustrations in the usual access-log format. A hit on the first route reads `203.0.113.7 - - [date] "GET /api/v1/auto_login HTTP/1.1" 200` and a hit on the second reads `"POST /api/v1/validate/code HTTP/1.1" 200`. A call to the first followed by the second from the same source, or from an address outside your network, is the sign to investigate. A 200 on auto_login from an outside address shows that a superuser token was issued.
- On the host, review child processes spawned by the Langflow Python process, new files under temporary directories, and outbound connections to unfamiliar hosts. KEV marks forensic triage as required for this CVE, so preserve logs and memory before you clean up.
- IBM, NVD and CISA publish no indicator addresses or file names for this CVE. Wiz Research's honeypot report lists these for attacks on its Langflow honeypot, without naming a CVE: staging server 94.26.106[.]29, a dropper at /tmp/x86_64 or /tmp/amd64 that deletes itself, and a miner binary renamed unicorn under /app/data/.claude/. Treat a match as a lead to investigate, not as proof of this CVE.

## Timeline

| Date | Event | Source |
| --- | --- | --- |
| 23 Jun 2026 | Langflow 1.10.1 is released. Its notes list 'do not execute code in validate_code' and 'stop issuing 365-day superuser token via auto_login'. | [source](https://github.com/langflow-ai/langflow/releases/tag/v1.10.1) |
| 2 Jul 2026 | IBM publishes the security bulletin (initial publication) and credits Harish Kolla. | [source](https://www.ibm.com/support/pages/node/7278927) |
| 17 Jul 2026 | NVD publishes CVE-2026-9198. The CISA coordinator SSVC entry on the record carries the same date, with exploitation active, automatable yes, technical impact total. This is CISA's assessment stamp and the earliest dated mark of exploitation, not an observed attack. No report names the attacker or the first attack. | [source](https://services.nvd.nist.gov/rest/json/cves/2.0?cveId=CVE-2026-9198) |
| 4 Aug 2026 | CISA adds the CVE to KEV with forensic triage required and a due date of 7 Aug 2026. Ransomware use is listed as Unknown. | [source](https://www.cisa.gov/known-exploited-vulnerabilities-catalog?field_cve=CVE-2026-9198) |

The earliest dated mark of exploitation is CISA's, on 17 Jul 2026, 15 days after disclosure; exploitation may have begun earlier.

## Exploitation reported so far

Wiz Research's report of 27 Aug 2026 covers 90 days of telemetry from honeypots that include Langflow and records attackers staging tools on the Langflow honeypot. It names no CVE and gives no date for those attacks, so it does not date the first exploitation of CVE-2026-9198.

## Why runtime reachability settles it

The default configuration of an affected version is vulnerable, so version alone does not decide exposure: the running instance must have auto-login on and be reachable by an outsider. Those are runtime facts, shown by the effective LANGFLOW_AUTO_LOGIN value and the listening address. KEV and CISA's SSVC entry confirm active attacks. See [runtime reachability](/methods/runtime-reachability/) for what the method proves and misses.

## Related pages

- [CVE-2026-53266: Linux kernel](/ledger/cve-2026-53266/)
- [CVE-2026-59822: BerriAI LiteLLM, LiteLLM MCP endpoint auth bypass](/ledger/cve-2026-59822/)
- [CVE-2026-46300: Linux Kernel](/ledger/cve-2026-46300/)

## Sources

- CVE-2026-9198: NVD record: https://services.nvd.nist.gov/rest/json/cves/2.0?cveId=CVE-2026-9198 (read 19 Nov 2026)
- CVE-2026-9198: CISA KEV entry (feed version 2026.11.18): https://www.cisa.gov/known-exploited-vulnerabilities-catalog?field_cve=CVE-2026-9198 (read 19 Nov 2026)
- CVE-2026-9198: vendor advisory (IBM security bulletin): https://www.ibm.com/support/pages/node/7278927 (read 19 Nov 2026)
- CVE-2026-9198: vendor release notes for 1.10.1: https://github.com/langflow-ai/langflow/releases/tag/v1.10.1 (read 19 Nov 2026)
- CVE-2026-9198: vendor documentation on authentication and LANGFLOW_AUTO_LOGIN: https://docs.langflow.org/api-keys-and-authentication (read 19 Nov 2026)
- CVE-2026-9198: GitHub advisory database record: https://api.github.com/advisories?cve_id=CVE-2026-9198 (read 19 Nov 2026)
- CVE-2026-9198: Wiz Research honeypot report: https://www.wiz.io/blog/ai-infrastructure-honeypot (read 19 Nov 2026)

Last reviewed 19 Nov 2026. Canonical: https://reachabilityanalysis.com/ledger/cve-2026-9198/
