Advisory

GHSA-r95q-fp26-h3hc

Details, part 2 of 2. The advisory

Details

POST /api/esri HTTP/1.1
Host: cloudtak.example.org
Content-Type: application/json

{ "url": "http://169.254.169.254/latest/meta-data/arcgis/rest" }

Returns 403 Authentication Required (from Auth.is_auth → api/lib/auth.ts:118). The SSRF is reachable by any authenticated user but not by an anonymous one.

E2E reproduction (sink path against a controlled internal victim)

Because exercising the real route against a public cloud metadata endpoint is not something to do against third-party infrastructure, the sink was reproduced against a local internal-only victim using the same `fetch` import (`@tak-ps/etl`) and the verbatim `EsriBase.sniff()` + `fetchVersion()` logic the route uses. The harness models the route handler exactly: it new URL()s the user input, runs sniff(), then fetch()s — with isSafeUrl deliberately *not* called (matching shipped behavior), and a toggled control branch that calls it (matching the basemap guard).

Victim (victim.mjs) — an internal-only service on 127.0.0.1:9099 returning a secret JSON body:

[victim] internal service listening on 127.0.0.1:9099
[victim] HIT GET /arcgis/rest?f=json from 127.0.0.1

Harness output (esri_ssrf_e2e.mjs), user-supplied URL http://127.0.0.1:9099/arcgis/rest:

[VULN/no-guard] route returned full internal body:
{
  "type": "SERVER",
  "base": "http://127.0.0.1:9099/arcgis/rest",
  "upstreamBody": {
    "currentVersion": "11.4",
    "internal-only": true,
    "aws-metadata-simulated": {
      "iam": { "role": "cloudtak-prod-instance-role", "AccessKeyId": "ASIA_FAKE_INTERNAL_KEY_DO_NOT_USE" }
    },
    "note": "If you can read this from a user-supplied URL, that is SSRF (full-read)."
  }
}
[CONTROL/guarded] blocked as expected: Blocked URL: blocked IP address: 127.0.0.1

And isSafeUrl confirms it *would* block the metadata / loopback targets if it were called on this path:

http://169.254.169.254/latest/meta-data/ => {"safe":false, ... "reason":"blocked IP address: 169.254.169.254"}
http://127.0.0.1:9999/                  => {"safe":false, ... "reason":"blocked IP address: 127.0.0.1"}
http://localhost/                       => {"safe":false, ... "reason":"blocked hostname: localhost"}

So: with the shipped (no-guard) code the request reaches the internal victim and the full internal body is returned; with the basemap-style isSafeUrl guard the same request is blocked. (A full container OOM-style demonstration of reading real cloud metadata is intentionally not performed against live infrastructure; the victim-host reproduction is the honest, self-contained equivalent of the route's fetch path.)

Root cause

Two compounding gaps:

  1. No IP/DNS classification before fetch. api/lib/esri.ts imports the unguarded fetch from @tak-ps/etl and calls it with a URL derived directly from user input in every EsriProxy* method and in EsriBase.fetchVersion() / generateToken(). Nothing resolves the hostname and rejects private / loopback / link-local addresses. The repository already depends on @tak-ps/node-safeurl (isSafeUrl) precisely for this, and uses it in api/routes/basemap.ts, api/routes/task.ts, and api/lib/control/video-service.ts — but the guard was never added to the ESRI route family or to the ESRI library. This is an incomplete migration: the SafeURL hardening (PR #1468) covered basemap/task/video but left /api/esri* and api/lib/esri.ts (including the ESRI sub-branch of basemap import) unprotected.
  1. `sniff()` validates the wrong thing. The only inspection the URL receives before being fetched is EsriBase.sniff(), which pattern-matches the *pathname* for /rest, /arcgis/rest, or /sharing/rest. It never looks at the host, so an attacker simply appends /rest (or /arcgis/rest) to an internal URL and it is accepted and fetched.

Impact

  • CWE-918 Server-Side Request Forgery, full-read.
  • Cloud credential theft: reading http://169.254.169.254/latest/meta-data/iam/security-credentials/... (IMDSv1) yields the instance role's temporary AWS credentials, which an attacker can use against the deployment's cloud account.
  • Internal network enumeration and data exfiltration: any host:port reachable from the CloudTAK server (loopback admin ports, VPC-internal services, databases with HTTP interfaces, link-local) can be probed and, where the response is JSON-ish, read in full via the reflected response body / error string.
  • Privilege required: any authenticated user with any token (anyResources: true) — not limited to administrators. Unauthenticated requests are rejected (403), so this requires a valid account but no special role.

Fix

Centralize the existing isSafeUrl guard inside the ESRI library so every /api/esri* route and the ESRI sub-branch of basemap import are covered by one chokepoint, mirroring the guard already used in basemap.ts:

  • Add an async URL-classification step that runs isSafeUrl(...) and throws Blocked URL: <reason> for any URL that resolves to a private/loopback/link-local address, before the first fetch in EsriBase (e.g. in EsriBase.from() and a guarded constructor/init path, so both new EsriBase(...) + later proxy calls and EsriBase.from(...) are covered).
  • Preserve the existing process.env.StackName !== 'test' test-mode skip used elsewhere so the test suite is unaffected.

Because all six routes funnel through EsriBase / the EsriProxy* classes in api/lib/esri.ts, guarding the library is sufficient and avoids re-introducing the same per-route omission. A reference patch implementing exactly this (guard added to the ESRI library + applied on the route entry points, matching the basemap pattern) is provided as a pull request from a private fork.

Disclosure

  • Reported by tonghuaroot.
  • Credit: tonghuaroot only.

---

Reference fix PR (private advisory fork): https://github.com/dfpc-coe/CloudTAK-ghsa-r95q-fp26-h3hc/pull/1 — commit ff6dd1d9, centralizing isSafeUrl inside api/lib/esri.ts (safeFetch wrapper + EsriBase.assertSafe).

Sources & licences