Agent workflow exports
Check an agent that is defined outside source code, in a workflow export from a platform such as n8n. You get findings, a gate and evidence files, like a scan, without running the workflow. (A compiled helper is checked on the binary analysis page.)
Before you start
Section titled “Before you start”- The Keryn CLI (install). Or Keryn Web: drop the file, or a folder or zip that contains it.
- An export from your workflow platform (JSON or YAML).
- Nothing is executed, fetched or sent. Keryn reads the file.
Risky combinations found in an export can be drawn and analysed as chains on the Trust Canvas.
1. Analyse a workflow export
Section titled “1. Analyse a workflow export”Here is a small n8n export, written for this page: a webhook receives a ticket, an agent node asks a model, and the model’s answer is run as a command before the result is posted to a URL.
{ "name": "Support triage", "nodes": [ {"id": "1", "name": "Ticket webhook", "type": "n8n-nodes-base.webhook", "typeVersion": 1, "position": [0, 0], "parameters": {"path": "ticket", "httpMethod": "POST"}}, {"id": "2", "name": "Ask model", "type": "@n8n/n8n-nodes-langchain.agent", "typeVersion": 1, "position": [200, 0], "parameters": {"text": "={{ $json.body.text }}"}}, {"id": "3", "name": "Run command", "type": "n8n-nodes-base.executeCommand", "typeVersion": 1, "position": [400, 0], "parameters": {"command": "={{ $json.output }}"}}, {"id": "4", "name": "Post result", "type": "n8n-nodes-base.httpRequest", "typeVersion": 4, "position": [600, 0], "parameters": {"url": "https://hooks.example.invalid/result", "method": "POST"}} ], "connections": { "Ticket webhook": {"main": [[{"node": "Ask model", "type": "main", "index": 0}]]}, "Ask model": {"main": [[{"node": "Run command", "type": "main", "index": 0}]]}, "Run command": {"main": [[{"node": "Post result", "type": "main", "index": 0}]]} }}keryn workflow analyze flow.json --no-confign8n export (explicit_graph) · 4 nodes · 4 tools · 8 dangerous compositions · 1 external effects · approval coverage 0% · 2 findingscoverage: n8n export (explicit_graph): 4 nodes, 4 tools, 8 dangerous compositions, 1 external effects, approval coverage 0%; 2 findings from 28 analyzers; no rules executed.result: FAIL (exit 1) - 2 violation(s)/finding(s) at or above 'error'Keryn normalised the export into a workflow graph, whose edges are the links between steps, and ran its analysers over it: tool authority, approval boundaries, blast radius, dangerous compositions (risky combinations of steps on one path, such as code execution and an external send) and a threat model. The platform was detected from the export’s shape; keryn workflow platforms lists the 18 it knows, and --platform forces one. The coverage line says what ran: the analysers, and no rules.
The two findings are the high-impact steps with no approval boundary. List them with their ids:
keryn workflow analyze flow.json --no-config --sarif - --quiet | jq -r '.runs[0].results[] | "\(.level) \(.message.text)"'error Code execution + network access (Ticket webhook, Ask model, Run command, Post result)error High-impact action path without enforced approval (Ticket webhook, Ask model, Run command, Post result)The same weakness is reported once. A composition risk, an approval gap and the threat-model entry that restates them are folded into one finding, and the folded ids are kept as correlated evidence.
2. Keep the evidence files
Section titled “2. Keep the evidence files”keryn workflow analyze flow.json --no-config --evidence-dir evidence --quiet || truels evidencemanifest.jsonseceng-workflow-analysis.jsonseceng-workflow-graph.mmdseceng-workflow-report.mdseceng-workflow-risk-register.csvseceng-workflow-threat-model.jsonseceng-workflow-xyflow.jsonworkflow.sarif--evidence-dir writes the Markdown report, a Mermaid diagram of the graph, a CSV risk register, the threat model, a graph file for diagram tools and SARIF. (File names keep the product’s earlier prefix; the content is the analysis above.)
3. Binaries and strings dumps
Section titled “3. Binaries and strings dumps”A compiled helper, an MCP server binary or a strings dump is checked with keryn artifact analyze, which has its own page: binary analysis. It works the same way as the workflow command above: static, nothing executed, findings that gate. Neither command produces an AI-BOM. In Keryn Web, both kinds of file found inside a scanned folder or zip are analysed automatically and join the result.
What you just proved
Section titled “What you just proved”An agent’s authority is not only in its code. A workflow export is a graph of calls and sends with no source file. You checked it without running it, and you got the same kind of answer as for code: which steps can act without an approval, what they reach, and how much of that Keryn could see. The limits are the usual ones: this is a static reading of an export, and a clean result is not a statement that the workflow is safe.
- Compiled helpers and strings dumps: binary analysis.
- Draw the chain and find its chokepoint: Trust Canvas.
- Gate it in the pipeline: CI.
- Read a flow in a repository scan: read an attack chain.