All articles

NewsOctober 30, 20264 min read

Langflow: critical flaws allow auth bypass and RCE

IBM fixes four vulnerabilities in Langflow 1.13.0, two of them critical (CVSS 9.8): access without authentication and code execution via MCP.

A developer at a keyboard surrounded by floating code windows and digital interfaces

IBM has released Langflow 1.13.0, which fixes four vulnerabilities in the open source framework for building artificial intelligence agents, workflows based on large language models (LLMs) and retrieval-augmented generation (RAG) systems. Two are critical: they make it possible to bypass authentication and to execute code on the server without credentials. On 30 October the Centre for Cybersecurity Belgium (CCB) published an advisory urging organisations to patch immediately; the story was then covered by The Hacker News and CSO Online.

Langflow, now owned by IBM, is popular in companies because it lets users build AI workflows and agents in a visual editor, without writing code. For exactly that reason its instances tend to hold API keys for model providers, database credentials and access to internal systems. According to CSO Online, earlier vulnerabilities in the framework have already been exploited to steal AI service and developer credentials, most recently in August.

The fixed vulnerabilities

The technical details that follow come from CSO Online, the only one of the three sources whose full text could be read:

  • CVE-2026-93680 – critical (CVSS 9.8). It stems from an incomplete fix of a previous issue related to automatic login (auto-login). An attacker can bypass security restrictions and use features that would normally require authentication.
  • CVE-2026-110186 – critical (CVSS 9.8). It allows an unauthenticated remote attacker to execute arbitrary code through the enumeration of Model Context Protocol (MCP) tools, the protocol AI agents use to discover and call external tools.
  • CVE-2026-110193 – high. Code and command execution on the server, caused by an issue in a third-party library used by the Agentic Assistant. Exploiting it requires an account, even a low-privileged one.
  • CVE-2026-110192 – medium (CVSS 6.5). Information disclosure: API credentials are not sufficiently protected.

Langflow OSS versions 1.0.0 to 1.12.8 are affected. Version 1.13.0 is available both from PyPI and as a Docker image.

Again according to CSO Online, IBM also fixed three medium-severity vulnerabilities in Hono, the API and web application framework Langflow uses for request routing and some web server tasks, and recommends upgrading it to version 4.13.5.

The sources consulted do not report active exploitation of these four vulnerabilities. The headline of The Hacker News, however, highlights the most sensitive point: unauthenticated attackers can run code on servers.

Why the risk is high

The two critical flaws add up: one opens restricted features without logging in, the other leads straight to code execution without credentials. On an instance reachable from the Internet the vector is the classic exploitation of a public-facing application (MITRE ATT&CK T1190). Once inside, an attacker finds on the server the API keys and credentials configured in the workflows (T1552, unsecured credentials), which can open the way to cloud services, databases and document repositories connected to RAG systems.

Impact for Italian and European companies

Langflow is often installed by development or data science teams for prototypes and pilot projects, sometimes outside normal IT management processes: some instances may well be missing from the inventory and remain unpatched. For entities in scope of NIS2, vulnerability handling and software supply chain security are among the required measures, and a forgotten AI component is fully part of that perimeter. Where RAG workflows process internal documents, including in healthcare, a compromise can turn into a personal data breach, with the notification obligations set by the GDPR.

What to do now

  1. Take inventory. Look for Langflow instances in production, in Docker containers and in development and test environments, including those installed with pip on workstations.
  2. Upgrade to Langflow 1.13.0. Before upgrading, IBM recommends backing up the database, the secret key and uploaded files. If the instance uses automatic login, prepare the existing account for password sign-in, as described in the installation and Docker upgrade instructions.
  3. Upgrade Hono to 4.13.5, as recommended by IBM.
  4. Take Langflow off the Internet. Until you upgrade, and as good practice in any case, make the interface reachable only from the internal network or through a VPN, with authentication in front of it.
  5. Rotate credentials. If a vulnerable instance was exposed, treat the API keys and credentials configured in the workflows as compromised and replace them.
  6. Check the logs. Look for unauthenticated requests to restricted features and MCP endpoints, unexpected child processes of the Langflow service and unusual outbound connections from the server.

How ARXDOME helps

ARX Expose identifies services exposed on the Internet and known vulnerabilities in recognised products, including a Langflow interface published by mistake. ARX Shield shows open vulnerabilities per machine and per package on monitored machines, highlighting those already exploited, while ARX Sentry keeps watch over security events on the servers running these services.

Last updated:

Any questions?

Let’s talk about your company’s needs, with no obligation.

Contact us

Related articles