Reported by @tg12 in three audit issues opened the same day:
#299 — POST /api/sentinel/token is an unauthenticated Copernicus
OAuth relay for caller-supplied client_id/secret.
#300 — POST /api/sentinel/tile is an unauthenticated quota/bandwidth
relay for Sentinel Hub Process API tile fetches.
#301 — GET /api/sentinel2/search is an unauthenticated Planetary
Computer STAC + Esri imagery search relay.
All three lived in backend/routers/tools.py decorated only with
@limiter.limit(...) — no Depends(require_local_operator). That made
the backend a free anonymous relay for any caller's Sentinel /
Planetary Computer queries, in the same shape we already closed for
#240/#241 (oracle resolve) and #211/#213/#214 (thermal verify, OpenMHZ
calls + audio relay).
Fix: add dependencies=[Depends(require_local_operator)] to each route.
Loopback / Docker-bridge / admin-key callers (the operator dashboard)
are unaffected — they still resolve through the same allowlist used by
every other operator-only helper in this file. Anonymous remote callers
now receive 403 BEFORE any outbound HTTP call to Copernicus or
Planetary Computer happens.
Tests
-----
test_sentinel_routes_auth_gate.py — 8 new tests:
* anonymous-remote → 403 on all three routes
* NO upstream HTTP call when the gate fires (asserted via
MagicMock(side_effect=AssertionError) on requests.post and
services.sentinel_search.search_sentinel2_scene). This is the
property that makes the gate real — without it, a 403 returned
after the upstream call still burns quota.
* 127.0.0.1 loopback caller reaches the handler (no false-positive
where the gate accidentally blocks the local operator too).
* Uses raw ASGITransport(client=(peer_ip, ...)) rather than
FastAPI's TestClient because TestClient reports client.host as
"testclient" which is not on the loopback allowlist.
test_control_surface_auth.py — extended the existing parameterised
regression with the three new routes. That regression is the global
"no remote control surface ships without auth" guard for the whole
codebase; adding these to it means a future refactor that drops the
dependency from any of them will fail CI alongside the existing
~30 gated routes.
The egress-on-403 property and the parameterised regression together
give two independent proofs that the gate fires before the upstream
network call, even if FastAPI's internal dependant tree shape changes
across versions (an earlier draft of this PR included a static walker
of the route table; it was removed because behavioural evidence is
strictly stronger and version-independent).
External security audit by @tg12 (May 17, 2026) filed 11 issues against
the backend. PR #227 (May 18, AI-generated) closed seven of them by
adding require_local_operator to control-plane endpoints. Four remained
live; this PR closes the rest.
#192 — CCTV proxy followed redirects without re-validating host
Issue: /api/cctv/media validated only the caller-supplied URL host
before passing it to requests.get(..., allow_redirects=True). A 302
to http://127.0.0.1 or any internal/disallowed host was silently
followed, turning the proxy into an open-redirect-to-SSRF chain.
Fix in routers/cctv.py: replace the single allow_redirects=True call
with a manual follow loop. Each hop's Location is parsed, the host is
rerun through _cctv_host_allowed(), and non-HTTP schemes (file://,
ftp://, etc.) are rejected. Cap chain length at 5 hops.
Test: backend/tests/test_cctv_redirect_ssrf.py covers
- redirect to disallowed host -> 502
- redirect to localhost -> 502
- redirect to another allowed host -> 200
- redirect chain length cap
- non-HTTP scheme rejected
#198 — Gate introspection GETs were unauthenticated
Issue: /api/wormhole/gate/{gate_id}/{identity,personas,key} were
callable with no auth dependency. Any caller that could reach the
backend could dump the operator's active persona, persona inventory,
and key status for any gate_id they knew. The wiki's privacy threat
model explicitly markets gate personas as rotating, unlinkable
pseudonyms — this leak defeated that property.
Fix in routers/wormhole.py: add
dependencies=[Depends(require_local_operator)] to all three routes.
Test: backend/tests/test_control_surface_auth.py extended with
three new parameterized cases (lines 75-77).
#199 — GDELT military incident ingestion used plaintext HTTP
Issue: backend/services/geopolitics.py fetched
http://data.gdeltproject.org/gdeltv2/lastupdate.txt and ~48 export
archive URLs over plaintext HTTP. Passive observers could identify
Shadowbroker nodes from the fetch pattern. Active MITM could inject
doctored military incident records into the global map.
Fix in services/geopolitics.py: rewrite the lastupdate.txt fetch and
the export download URL constructor to use https://. GDELT's
data.gdeltproject.org serves the same content over HTTPS.
Test: backend/tests/test_gdelt_https.py asserts no plaintext HTTP
URLs to data.gdeltproject.org remain in code (comments excluded) and
that the HTTPS URLs we expect are present.
#200 — Sentinel token cache lookup used client_id only
Issue: routers/tools.py kept a process-global cache of Copernicus
bearer tokens. The lookup compared
_sh_token_cache["client_id"] == client_id. A caller who knew a valid
client_id but supplied any wrong client_secret hit the cache and
reused the legitimate caller's bearer token — burning their quota
and accessing imagery on their account.
Fix in routers/tools.py: replace the client_id field with
credential_fp, an HMAC-SHA256 over (client_id, client_secret) under
a per-process random key (_SH_TOKEN_CACHE_HMAC_KEY = os.urandom(32),
regenerated at startup). A caller who doesn't know the secret cannot
compute a matching fingerprint, so they miss the cache and hit the
real Copernicus token endpoint — which will reject their wrong
secret with a 401.
Test: backend/tests/test_sentinel_token_cache.py covers
- same client_id + different secrets => different fingerprints
- same credentials => same fingerprint (cache still works)
- different client_ids + same secret => different fingerprints
- cache no longer stores raw client_id (catches regression)
- attacker with wrong secret cannot reuse victim's token
Validation
pytest backend/tests/test_control_surface_auth.py
backend/tests/test_cctv_redirect_ssrf.py
backend/tests/test_gdelt_https.py
backend/tests/test_sentinel_token_cache.py
-> 37 passed
Credit: @tg12 reported all four of these in their May 17 audit with
correct line-number citations and accurate remediation recommendations.
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
BigBodyCobain