FlowSentry

Security scanner for n8n workflows — 18 rules, zero dependencies, CI-ready.

FlowSentry parses your n8n workflow JSON exports and flags the failures we keep seeing in the wild: unauthenticated webhooks, hardcoded secrets, SSRF to cloud metadata, shell injection through Execute Command, eval-style code nodes, SQL built from user input, over-scoped credentials and more. Output: terminal, JSON, SARIF 2.1.0 (GitHub Code Scanning) and a self-contained HTML report.

Zero runtime dependencies (pure Python stdlib, Python 3.9+)
18 rules mapped to the OWASP Agentic Top 10 (2026)
SARIF output → upload straight to GitHub Code Scanning
Exit code 1 on findings → drop it into CI as a quality gate

bash

$ pip install flowsentry
$ flowsentry scan workflows/ --sarif flowsentry.sarif

PyPI: https://pypi.org/project/flowsentry/ · or install straight from GitHub: pip install git+https://github.com/vasilicasijarvis/flowsentry.git

No install at all:

bash

git clone https://github.com/vasilicasijarvis/flowsentry
cd flowsentry
python3 flowsentry_cli.py scan ./workflows

Why

n8n webhooks are unauthenticated by default. Most self-hosted instances sit behind a single reverse proxy, and one exported workflow is enough to leak a hardcoded API key or an Execute Command node that interpolates request data straight into a shell.

We scanned 10 real, public n8n workflow files from GitHub with FlowSentry v0.1:

File (source repo)	Critical	Medium	Notable finding
`AI_Bot.json` (AnaamRasool/WhatsApp-Bot)	1	4	WhatsApp webhook with `authentication: none`, replies echo full node output
`00485-library-install.json` (DragonJAR/n8n-workflows-esp)	2	1	Bash script interpolating `{{$json.library}}` into shell
`n8n/w1.json` (Finfra/dockers)	1	2	Webhook `authentication: 'none'`
`workflows/103.json` (n8n-io/test-workflows)	4	2	`echo 'test' > /tmp/{{$node["Set"].json["filename"]}}`
`subagente-citas.json` (santifer/jacobo-workflows)	1	9	Public appointment-booking webhook, no auth
`nl2sql.json` (Wilsonoonn/n8n_nl2sql)	1	9	Natural-language-to-SQL behind an unauthenticated webhook
`api-authentication.json` (yorrickjansen/n8n-webhook-security)	1	5	Demo webhooks themselves accept unauthenticated requests
`AIDA Preisalarm`, `nl2sql` helper, `openwebui-pipe` (3 more files)	0	7	code nodes, missing error handling, response modes

Totals: 11 critical, 39 medium across 10 files. 100% of files had at least one finding.

These are public workflow exports shared by their authors for learning/demo purposes — nothing was exploited and no instance was accessed. FlowSentry is defensive tooling: it reads JSON files, it does not send requests.

Real findings, verbatim

1. Shell injection via Execute Command (n8n official test workflow):

Code

[!] CRITICAL FS005 Execute Command node without guardrails
    node: Execute Command  (n8n-nodes-base.executeCommand)
[!] CRITICAL FS009 Expression-based command injection
    node: Execute Command
    evidence: echo 'test' > /tmp/{{$node["Set"].json["filename"]}}
    fix: Never interpolate data into commands; use argv-style APIs or strict
         allowlist validation of the entire command string.

2. Unauthenticated webhooks (5 of 10 files):

Code

[!] CRITICAL FS001 Webhook endpoint without authentication
    node: HTTP Trigger  (n8n-nodes-base.webhook)
    Webhook node has authentication set to 'none'. Anyone who can reach the
    n8n instance can trigger this workflow and its downstream actions.
    fix: Set Webhook > Authentication to Basic/Header/JWT auth, or validate a
         shared-secret header in the workflow before doing anything sensitive.

3. What the other rules catch (crafted example from the test suite):

Code

[!] CRITICAL FS002 Hardcoded secret in node parameters
    Parameter 'accessToken' looks like it contains a hardcoded secret
    (literal value, no expression). Exports of workflows leak like this.
    fix: Move the value into an n8n credential and reference it via expressions.

Rules FS002–FS004, FS011–FS013 and FS015–FS018 (secrets, SSRF/IMDS, SQL injection, plain http, community nodes, exfil sinks) did not fire on this particular sample but are fully covered by the 40-test suite in tests/test_rules.py.

Real run, 2026-09-06 — flowsentry scan examples/real on the 10 public workflow exports (exit code 1 as expected on critical findings):

Code

  examples/real/AI_Bot.json  (WhatsApp bot) - 5 finding(s)
    [!] CRITICAL FS001 Webhook endpoint without authentication
           node: HTTP Trigger  (n8n-nodes-base.webhook)
    [~] MEDIUM   FS017 Webhook response mode exposes internals
           node: API: JWT auth with auth server validation  (n8n-nodes-base.webhook)
           Webhook responseMode='responseNode' - the last node's full output (possibly
           including credentials, internal IDs, stack traces) is returned to the
           unauthenticated caller.
           fix: Return an explicit minimal payload via the Respond to Webhook node.

  Summary
    critical: 11   high: 0   medium: 39   low: 0
    Result: FAIL - fix critical/high findings before production.

Full machine-readable results: examples/scan_report.json, examples/scan_report.sarif and examples/scan_report.html.

Rules (v0.1)

Rule	Severity	Detects
FS001	critical	Webhook endpoint without authentication
FS002	critical	Hardcoded secret in node parameters
FS003	high	Hardcoded secret in HTTP header/query
FS004	critical	SSRF / cloud metadata endpoint access (IMDS 169.254.169.254, GCP, Alibaba)
FS005	critical	Execute Command node without guardrails
FS006	medium	Code node without sandbox hardening (no task runners)
FS007	critical	Dynamic code construction (eval, new Function, child_process, os, subprocess, dynamic `$()`)
FS008	medium	Credential over-scoping (admin/root/owner names, unusual types)
FS009	critical	Expression-based command injection (`{{$json...}}` into shell)
FS010	medium	Missing error handling (no errorWorkflow, no Error Trigger)
FS011	high	Raw SQL built from expressions
FS012	high	HTTP node over plain `http://`
FS013	medium	Exposed trigger (Form/Telegram/IMAP accepting unauthenticated input)
FS014	medium	Credential reuse across 5+ nodes
FS015	medium	Community/unknown node packages
FS016	medium	Data sent to exfil-style sinks (webhook.site, pastebin, ngrok…)
FS017	medium	Webhook response mode echoing internal data
FS018	medium	Set node storing secrets in plaintext

Usage

bash

# Scan files or directories (directories are walked for *.json)
flowsentry scan ./workflows
flowsentry scan export1.json export2.json

# Machine-readable outputs
flowsentry scan ./workflows --json report.json --sarif report.sarif --html report.html

# CI gate: exit 1 when findings at/above severity exist (default: high)
flowsentry scan ./workflows --fail-on critical
flowsentry scan ./workflows --fail-on never   # always exit 0

# List rules
flowsentry rules

GitHub Actions

yaml

name: flowsentry
on: [push, pull_request]
jobs:
  scan:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-python@v5
        with: { python-version: "3.12" }
      - run: pip install git+https://github.com/vasilicasijarvis/flowsentry.git
      - run: flowsentry scan ./workflows --sarif flowsentry.sarif --fail-on high
      - uses: github/codeql-action/upload-sarif@v3
        if: always()
        with:
          sarif_file: flowsentry.sarif

Scan an n8n export bundle

bash

# From your n8n UI: Workflows > Export (or via n8n CLI)
n8n export:workflow --all --output=./workflows
flowsentry scan ./workflows --html report.html

Demo

FlowSentry demo

Development

bash

git clone https://github.com/vasilicasijarvis/flowsentry
cd flowsentry
python3 tests/run_tests.py     # 40 tests, zero dependencies

Re-fetch the public example workflows used in the README scan:

bash

python3 scripts/fetch_workflows.py
flowsentry scan examples/real --json examples/scan_report.json

Roadmap

v0.2 — live scanning via the n8n REST API, drift detection (workflow changed since last scan)
v1.0 — FlowSentry Cloud: continuous monitoring, alerting, multi-instance dashboard
v1.5 — MCP server config auditing (tool poisoning, auth gaps, unpinned versions)
v2.0 — compliance evidence packs (OWASP Agentic Top 10, EU AI Act, SOC 2)

The hosted, continuous version is in the works — join the early list: flowsentry.vercel.app

License

Apache-2.0. Scan your own workflows or exports you have permission to analyze.