
Flowise is a popular open-source, drag-and-drop builder for LLM apps and AI agents. Its maximum-severity code-injection flaw, CVE-2025-59528, is now being exploited in the wild. The bug was disclosed in September 2025, and 12,000 to 15,000 Flowise instances are reportedly exposed to the internet. If your team spun up Flowise for an internal chatbot or agent prototype and never removed it from the internet, check it today.
What happened
CVE-2025-59528 was published on 22 September 2025. It carries a CVSS 3.1 base score of 10.0 from GitHub as the CNA (AV:N/AC:L/PR:N/UI:N/S:C/C:H/I:H/A:H) and is classed as CWE-94 code injection. The bug is in the CustomMCP node, which connects Flowise to an external Model Context Protocol server. The node parses the user-supplied mcpServerConfig string by passing it to JavaScript's Function() constructor without validating it first. Whatever is in that string runs with the full privileges of the Node.js process, including access to child_process and fs.
On 7 April 2026, VulnCheck researcher Caitlin Condon said the company's Canary network had detected the first exploitation of the flaw. The activity was limited and came from a single Starlink IP address. Condon also noted active exploitation of two older Flowise bugs, CVE-2025-8943 and CVE-2025-26319. VulnCheck shares its exploit samples, network signatures and YARA rules only with customers.
Who is affected
- Flowise 3.0.5 is the version listed in the GitHub/NVD record. Version 3.0.6 fixes the flaw. At the time of the report, the latest release was 3.1.1.
- Instances reachable from the internet are at highest risk, especially those without authentication in front of the UI and API.
- Teams that use Flowise as a backend for chatbots, RAG pipelines or agents. Those deployments often hold LLM provider API keys, vector database credentials and connector secrets, all of which a code-execution bug exposes.
What it means for developers and engineering teams
The root cause is a classic one: treating configuration as code. Config for an MCP server should be data, parsed with JSON.parse and checked against a schema. Running it through Function() or eval makes every form field that feeds it a remote shell. Low-code AI builders make this risk worse. Their job is to let users wire up arbitrary integrations, so the line between "configure a node" and "run code on the server" is thin. Several nodes across this class of tools have had similar bugs.
The impact goes beyond the Flowise host. An agent platform usually holds the keys to everything the agent can reach: OpenAI or Anthropic API keys, database URLs, SaaS OAuth tokens and internal APIs. Once an attacker has code execution, they can read environment variables and the Flowise credential store, and then move into those systems. Treat Flowise, n8n, Langflow and similar tools as privileged infrastructure. Don't treat them as a prototype you can leave running.
MCP also widens the attack surface. Every new connector type is new parsing code that handles attacker-influenced input. When you adopt MCP clients, review how each one loads server definitions, and don't let untrusted users define them.
What to do now
- Upgrade. Move to the current Flowise release (3.1.1 or later at the time of reporting, and never older than 3.0.6). For Docker deployments:
docker pull flowiseai/flowise:latest docker compose up -d --force-recreate - Take it off the public internet if nobody outside needs it. Put it behind a VPN, an identity-aware proxy or at least authentication at the reverse proxy.
- Assume compromise if it was exposed and unpatched. Look for unexpected child processes of the Node process, new files in the container, and outbound connections to unknown hosts.
- Rotate secrets stored in Flowise or its environment: LLM API keys, database credentials and any OAuth tokens for connected tools.
- Check the other exploited CVEs (CVE-2025-8943 and CVE-2025-26319) against your version, and patch them too.
- Restrict who can create or edit flows, since anyone who can configure a CustomMCP node on a vulnerable version can run code.