Skip to content

Outbound URL checks ​

Nodes that fetch a URL the workflow author or upstream data supplies — HTTP Request, Web Browse, the Chart node's data.url, the PDF node in HTML-render mode (including every sub-resource and WebSocket the page opens), and RAG endpoints set on a node (a vector store's endpoint, an embedder's baseUrl, the Azure OpenAI embeddings node's endpoint), plus the RSS trigger's feed URL and the SSE trigger's stream URL — run that URL through a shared check before any request leaves the host. This blocks server-side request forgery (SSRF) and local-file reads through those nodes.

Concepts: the lethal trifecta shows how URLs become a way for data to leave an agent.

MCP HTTP servers do not go through these checks. For the OAuth discovery, registration, token and refresh requests of remote MCP servers, set MCP_OAUTH_BLOCK_PRIVATE_HOSTS=true (off by default) to refuse private, loopback, link-local and cloud metadata addresses; see OAuth for remote servers. Limit where the API host can connect to MCP servers with egress rules at the network layer (a firewall or an outbound proxy).

Policy ​

  • Schemes. Only http: and https: can be fetched. file:, data:, javascript: and anything else are refused, so one of these nodes can never read the local disk or run inline script through a URL.
  • Always blocked. Cloud metadata endpoints (169.254.169.254, fd00:ec2::254, metadata.google.internal and the like) and link-local, multicast, unspecified, reserved and documentation ranges. The allowlist never overrides these.
  • Private ranges and localhost are allowed by default. Turn on OUTBOUND_DENY_PRIVATE to block them, then use OUTBOUND_ALLOWLIST to let specific internal hosts or networks back through.

The host is resolved and every address it resolves to is classified, so a name that points at a blocked address is refused. Redirects are re-checked hop by hop. A redirect to another origin drops the credential headers Authorization, Proxy-Authorization, Cookie, api-key and x-api-key, and the browser-based nodes check each sub-resource the page requests, not just the first document. WebSockets a page opens (ws: / wss:) get the same host check; a refused socket is closed before it connects. A refused URL fails the node at once and is not retried.

Inline chart images. A Chart spec may use data:image/... URLs for image marks; they are decoded locally and never fetched. Each one is capped at 2 MB (encoded). Other data: URLs and file: loads are refused.

Credentials and node-supplied RAG endpoints. When a RAG node names its own endpoint or baseUrl, only the node's own secret (embeddingSecretRef) is sent there. The deployment's environment keys (OPENAI_API_KEY, QDRANT_API_KEY and the like) only ever go to the deployment's own endpoints, so a workflow author cannot point a node at their own server to collect them.

DNS rebinding. For HTTP Request, the Chart node, node-configured RAG endpoints and the RSS and SSE triggers, the connection is pinned. The socket's own DNS lookup goes through the check: the name is resolved once at connect time, every address is checked, and the socket can only connect to an address that passed. A name that answers a public address to the first check and a metadata or private address a moment later is refused. The URL keeps its hostname, so TLS SNI, certificate validation and the Host header are unchanged.

Configuration ​

VariableDefaultPurpose
OUTBOUND_DENY_PRIVATEfalseAlso block private ranges and localhost for the checked nodes.
OUTBOUND_ALLOWLISTemptyComma-separated hosts or CIDRs allowed through under OUTBOUND_DENY_PRIVATE, e.g. 10.0.0.0/8,*.internal,db.corp:5432.

What is not covered ​

  • HTTP proxies. The pinned fetches connect directly and ignore HTTP_PROXY, HTTPS_PROXY and NODE_USE_ENV_PROXY. On a host whose only way out is through a proxy, HTTP Request, Chart data.url, node RAG endpoints and the RSS/SSE triggers cannot reach the internet. Proxy support is not implemented: with a proxy, the proxy resolves the name, so the connection could not be pinned to the vetted address.
  • Browser-based nodes are not pinned. Web Browse and the HTML-render PDF path check each request and WebSocket at route time, but Chromium then resolves the name itself. A short-TTL rebinding name can still pass the check and connect elsewhere. Under OUTBOUND_DENY_PRIVATE that is a gap for private ranges; the metadata block is affected the same way.
  • Python code nodes on the host open their own sockets and never go through these checks. See Code execution.
  • Other browser channels. WebRTC, and requests a service worker makes, are not routed through the checks.
  • Deployment-level endpoints are infrastructure settings and are not checked: LLM provider base URLs, connector SDK endpoints, and RAG endpoints that come from environment variables (QDRANT_ENDPOINT, CHROMA_ENDPOINT and the like) rather than from a node. MCP HTTP servers and MCP OAuth discovery are not checked either (see above).

This website (not the L2M product) uses Google Analytics, only after you accept cookies. Cookie settings