Testing & Pitfalls¶
What it is¶
Authorization bugs rarely show up in functional tests — the happy path works, because the tester is using their own data. You have to test the unhappy paths deliberately: the wrong user, the wrong tenant, the wrong role, the wrong verb.
How it works¶
The authorization test matrix¶
For every endpoint, enumerate actors × actions × objects and assert the expected outcome:
| Actor | GET /invoices/{own} |
GET /invoices/{other user} |
GET /invoices/{other tenant} |
DELETE /invoices/{own} |
|---|---|---|---|---|
| Anonymous | 401 | 401 | 401 | 401 |
| Member | 200 | 404 | 404 | 403 |
| Tenant admin | 200 | 200 | 404 | 204 |
Token missing invoices:read |
403 | 403 | 403 | 403 |
The matrix is the specification. Generate tests from it so a new endpoint can't ship without rows.
The two-user test¶
The single most effective BOLA test: create resources as User A, then replay every request with User B's token. Anything other than 403/404 is a finding.
Editable source: docs/diagrams/two-user-test.drawio
Reference implementation¶
import pytest
CASES = [
# (actor, method, path_template, expected)
("anonymous", "GET", "/invoices/{alice_inv}", 401),
("alice", "GET", "/invoices/{alice_inv}", 200),
("bob", "GET", "/invoices/{alice_inv}", 404), # BOLA
("mallory_other", "GET", "/invoices/{alice_inv}", 404), # cross-tenant
("alice", "DELETE", "/invoices/{alice_inv}", 403), # BFLA
("acme_admin", "DELETE", "/invoices/{alice_inv}", 204),
]
@pytest.mark.parametrize("actor,method,path,expected", CASES)
def test_invoice_authorization(client, tokens, fixtures, actor, method, path, expected):
headers = {} if actor == "anonymous" else {"Authorization": f"Bearer {tokens[actor]}"}
r = client.request(method, path.format(**fixtures), headers=headers)
assert r.status_code == expected
@SpringBootTest
@AutoConfigureMockMvc
class InvoiceAuthorizationTest {
@Autowired MockMvc mvc;
static Stream<Arguments> cases() {
return Stream.of(
Arguments.of("bob", "SCOPE_invoices:read", HttpStatus.NOT_FOUND), // BOLA
Arguments.of("alice", "SCOPE_invoices:read", HttpStatus.OK),
Arguments.of("alice", "SCOPE_profile", HttpStatus.FORBIDDEN)); // scope
}
@ParameterizedTest
@MethodSource("cases")
void getInvoice(String sub, String authority, HttpStatus expected) throws Exception {
mvc.perform(get("/invoices/{id}", ALICE_INVOICE_ID)
.with(jwt().jwt(j -> j.subject(sub))
.authorities(new SimpleGrantedAuthority(authority))))
.andExpect(status().is(expected.value()));
}
}
package httpapi.authz_test
import data.httpapi.authz
test_owner_can_read if {
authz.allow with input as {"action": "read",
"subject": {"id": "alice"}, "resource": {"owner": "alice"}}
}
test_other_user_cannot_read if {
not authz.allow with input as {"action": "read",
"subject": {"id": "bob"}, "resource": {"owner": "alice"}}
}
Testing techniques¶
- Route inventory diff. Compare the router's registered routes (or the OpenAPI spec) against routes that declare an authorization rule. Fail CI on any route without one.
- Verb tampering. For every
GETthat works, tryPUT,PATCH,DELETE, andHEADwith the same token. - Parameter pollution. Send the ID in path and body (
/users/1with{"id": 2}), duplicate query parameters, and alternate encodings. - Property fuzzing. Add privileged fields (
role,is_admin,tenant_id,owner_id) to every write payload and assert they're ignored or rejected. - Response diffing. Compare responses across roles to catch fields that leak to lower-privileged callers.
- DAST and proxies. Tools like Burp Suite's Autorize extension or OWASP ZAP replay a high-privilege session's traffic with a low-privilege token — on systems you are authorized to test.
Common pitfalls¶
| Pitfall | Why it fails | Fix |
|---|---|---|
| Client-side authorization | Hidden buttons and routes don't stop direct API calls | Every check enforced server-side |
| Fail-open defaults | Exceptions or unknown roles fall through to "allow" | Explicit deny-by-default; test the unknown-role case |
| Trusting client identity | user_id, tenant_id, or role read from the request |
Identity and tenant come only from the verified credential |
| Check-then-act races (TOCTOU) | Permission checked, then object changes owner before the write | Check and write in one transaction / conditional update |
| Inconsistent checks across channels | REST checks ownership; GraphQL, gRPC, websocket, or export doesn't | Shared authorization layer used by every entry point |
| Stale permissions | Removed users keep access via long-lived tokens or caches | Short TTLs, cache invalidation on permission change |
| Error messages leak existence | 403 for "exists but not yours", 404 for "doesn't exist" | Return 404 for both when existence is sensitive |
| Only testing as admin | Developers test with full permissions, so nothing ever fails | Default test fixtures use the least-privileged role |
Checklist¶
- An authorization matrix exists for every resource and is turned into tests
- Two-user (BOLA) and cross-tenant tests run in CI
- CI fails on any route without an authorization rule
- Write endpoints are fuzzed with privileged properties
- Policies have unit tests covering both allow and deny
- Periodic manual or DAST testing with low-privilege tokens