# LangFlow: two 9.8 vulnerabilities open the door to your AI agents

> Two 9.8 vulnerabilities in Langflow allow unauthenticated code execution through MCP. They are only fixed in 1.13.0 and a public attack module is already available.
> By José Maldonado · Cybersecurity · 2026-11-03

LangFlow, one of the most widely used open source tools for building agents and workflows with [artificial intelligence](https://observatorioblockchain.com/en/ai/gemini-4-carbon-ciberseguridad-fairwind/), it has accumulated two critical vulnerabilities that allow an attacker without an account to execute code on the server. They are [CVE-2026-93680](https://www.ibm.com/support/pages/security-bulletin-langflow-oss-affected-multiple-vulnerabilities-2) and CVE-2026-110186, both with a score of 9.8 out of ten, and they affect all versions released between 1.0.0 and 1.12.8. The only way out is 1.13.0, released this week.

The problem is not just the severity, it is the surface. A LangFlow deployment is not an isolated application: it is the glue that connects models, tools and data, and it usually holds credentials for several services at once. Whoever executes code there does not get an ordinary machine; they get API keys, access to databases and the ability to manipulate the responses that other systems take as valid.

The specific vector that makes LangFlow an interesting target is the protocol it uses to offer its flows as tools to other AI applications. That channel, designed so that an agent can find and use third-party capabilities, allows listing what the server exposes and, in the worst case, even executing operating system commands without having presented a single credential.

## Two 9.8s and a single version that fixes them

The difference between the two vulnerabilities matters little to anyone who has to decide this afternoon whether to update. One allows authentication to be bypassed and the other opens up code execution through component generation, and both add to a third problem of poorly protected credential exposure. Together they paint a scenario in which the attacker gets in, executes and takes whatever they find.

The affected version ranges are wide. Any installation that has not kept pace with releases since 1.0.0 carries the problem, and in agent projects it is common to stay on a specific version because flows change behavior between updates. That habit, reasonable in a lab, becomes dangerous when the same server serves real users.

1.13.0 fixes it, but requires migrating. Anyone with custom flows or their own components should test them before pushing the change to production, which takes time just when the clock is running. It is the classic dilemma between updating fast and updating well, except that here the exploit is already written.

## MCP, the door nobody was watching

The Model Context Protocol has become the de facto standard for models to use external tools, and with it came a new surface that many companies connected without inventorying it. An MCP server that exposes actions on the real world needs authentication, limits and activity logging, and in practice many deployments were set up for a demo and stayed in production.

LangFlow offers its flows as MCP tools, which multiplies the interest in attacking it. If you can list those tools and one of them executes code or queries sensitive data, you have a direct path without needing to understand the business logic. Tool enumeration stops being a minor reconnaissance step and becomes the vulnerability itself.

There is already a [public penetration testing module](https://thehackernews.com/2026/11/critical-langflow-flaws-could-let.html) which connects to the MCP endpoint without credentials, opens a session and lists the available tools. The module does not execute code, but it demonstrates how easy it is to confirm whether a specific server is vulnerable. With that information, the rest is a matter of patience and of which tools each installation exposes.

## Why is it not in KEV and yet still urgent?

Neither of the two flaws appears in the [catalog of exploited vulnerabilities](https://www.cisa.gov/known-exploited-vulnerabilities-catalog) of the US agency, although LangFlow already has [six entries](https://www.cisa.gov/known-exploited-vulnerabilities-catalog) there, the latest in August with CVE-2026-9198. The absence of an entry does not mean calm: the catalog reflects confirmed exploitation, and that confirmation arrives late, especially in niche software.

The tool's history works against any sense of calm. Six previously exploited flaws indicate that there is real interest in attacking it, and a product that accumulates that track record tends to be in the crosshairs of anyone looking for cheap initial access to AI environments. The likelihood that the seventh joins the list is not speculative.

There is also a structural reason. Teams that build agents tend to prioritize making them work, not making them resist, and the pressure to deliver quick demos leaves open configurations that nobody reviews afterwards. The software flaw is half the problem; the other half is the deployment.

## How to check whether your installation is exposed

The first check is the version. If the server runs any build earlier than 1.13.0, it is affected by both critical flaws, without exceptions. The second is exposure, and it is the one that usually goes unnoticed: looking for open MCP endpoints, administration panels accessible from outside and services that respond without asking for credentials.

The third is content. A LangFlow deployment usually stores API keys from several providers, database connection strings and credentials for internal tools. If the server was exposed before updating, all those credentials must be rotated, not just the main model's.

It is also worth reviewing the access log for recent weeks, looking for unauthenticated sessions, calls to the tools endpoint and repeated errors in component enumeration. An attacker probing tools leaves a fairly recognizable trail if someone keeps the logs long enough.

And there is one measure that does not depend on LangFlow. Servers that execute actions on the real world should not live on the same network as customer databases. Segmenting that layer reduces the damage of any flaw, present or future, without needing to correctly predict which one will come next.

## The missing inventory

The question that should trigger a review is not whether your company uses LangFlow, but how many servers capable of executing actions are loose on the internal network. Platform teams have spent years inventorying databases and web services, and most do not have an equivalent list of AI tools connected to sensitive data.

Updating to 1.13.0, rotating the credentials that have passed through those servers and reviewing which MCP tools are exposed is the minimum. Then comes something harder: deciding what an agent can do without supervision. A flow that creates users, moves files or queries customer information deserves the same treatment as an account with privileges, because in practice it is one.

The lesson of this case is uncomfortable for the current enthusiasm. Building agents is easy, connecting them to everything is too, and security has been left at the step nobody wants to take because it slows down the demo. With two unpatched 9.8s on the market, the bill for that debt is starting to arrive.

What comes next does not depend on LangFlow, it depends on the sector. Every orchestration tool that becomes popular will repeat the pattern: first adoption, then the warnings. If we learned anything from application servers fifteen years ago, it is that the middle layer is where risk accumulates, and now that middle layer talks to models.
