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

IncidentAI infrastructure

Langflow: a logged-in user's MCP server command runs as a shell, and honeypots have counted 2,300 exploit attempts since September

CriticalLangflowPrevidian

CVE-2026-105740 is an authenticated remote code execution flaw in Langflow, rated CVSS 9.9 in the GitHub advisory GHSA-7w94-79vh-5mr2 that Langflow published on September 10, 2026: any logged-in user can run operating system commands on the server, because the command of an MCP server set up with stdio transport is passed to bash as typed and runs when the server list is fetched. The advisory lists 1.8.3 as affected and 1.9.0 as patched. A second flaw of the same kind, CVE-2026-105697, also rated 9.9, let the stdio transport launch whatever command and arguments a user put in an MCP server configuration before 1.10.3. Previdian's alert of October 5 marks the flaw as exploited from October 11 and shows its honeypot telemetry: the first attempt on September 19, 333 attempts on October 11, around 300 a day through the week after, and 2,310 attempts from 30 IP addresses in 17 countries by October 18; Previdian's page shows no listing in the known-exploited catalog yet. The attackers and their goals are unknown. Mitigations: upgrade past 1.10.3, keep auto-login off and passwords strong, review the MCP server configurations stored in every project, and keep Langflow off untrusted networks.

Replay the attack step by step in the Learning Center

How it happened

  1. An MCP server's command is run through bash as typed, so a logged-in user's configuration is a shell command on the server.
  2. The command runs whenever the server list is loaded, and stays in the project settings.
  3. Scanners began exploiting it in volume from October 11, three weeks after the first attempt against the honeypots.

Impact

Honeypots only, by Previdian's account; any exposed Langflow instance on which an attacker can log in is being scanned and exploited in the same way.

The fix

Upgrade to 1.9.0 or later (1.10.3 or later for CVE-2026-105697), disable auto-login, set strong unique passwords, limit accounts to trusted users, and keep Langflow behind authentication rather than on the internet.

Lessons

  • A field that becomes a shell command makes every account an administrator; the login is the only barrier left.
  • Exploitation in volume began three weeks after the first probe; the window between a disclosure and mass scanning is the time you have to patch.

Sources