Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

IncidentAI infrastructure

Langflow: IBM lists more than two dozen flaws, two of them code execution with no login, and an LLM-written lambda that runs on the server

CriticalIBMLangflow

IBM's security bulletin of October 2, "Langflow OSS is affected by multiple vulnerabilities", covers Langflow OSS 1.0.0 through 1.12.2 and lists more than two dozen CVEs, fixed in 1.12.3. Two are rated 9.8 and need no login: remote code execution through improper control of code generation (CVE-2026-104334) and through OS command injection (CVE-2026-93674). Most of the rest let a logged-in user run code, read or write files or reach stored credentials, among them a bypass of Langflow's own code security scanner and a dependency confusion flaw; the CVE records reached GitHub's advisory database on October 7. Separately, GitHub advisories published October 5 and 6 describe Smart Transform (CVE-2026-7700, CVSS 8.8, fixed in 1.10.3), which asks an LLM to write a Python lambda from the flow's instructions and its input and then runs it with full builtins, so content that steers the model becomes code on the server; the Python interpreter components that ran code unsandboxed (CVE-2026-10561, CVSS 9.9); and a credential encryption key derived from a predictable random seed (CVE-2026-9205, CVSS 9.1).

Replay the attack step by step in the Learning Center

How it happened

  1. Two flaws let anyone who can reach the server run code without logging in (code generation and OS command injection).
  2. Smart Transform evaluates the lambda an LLM writes after only a syntax check, so a prompt injection in the data it reads can become code execution.
  3. Stored credentials were encrypted with a key anyone could rebuild once they knew a short SECRET_KEY.

Impact

Code execution on Langflow servers, with and without a login, and exposure of the provider keys and database passwords Langflow stores.

The fix

Upgrade to Langflow 1.12.3, rotate the credentials stored in Langflow, keep it off the public internet, and limit who can create or edit flows.

Lessons

  • Code an LLM writes is untrusted input; running it needs a sandbox, not a format check.
  • A visual AI builder holds every key its flows use, so a flaw in it is a flaw in all of them.

Sources