[e2e] guardrail blocks the unselected model for path-routed providers

Add an always-on (no-credentials) e2e regression guard for the report
that a model-allowlist guardrail attached to a policy has no effect for
PATH-ROUTED providers, where the model travels in the URL rather than the
JSON body: Google Vertex (…/models/{model}:rawPredict) and AWS Bedrock
(/model/{id}/invoke).

Each provider is tested in isolation: its own provider, its own guardrail
whose allowlist holds only that provider's allowed model, and its own
policy — so the guardrail the proxy enforces is never a mixed
cross-provider list. Over the tunnel, the allowed model (in the URL path)
is served 200 and an unselected model is denied 403 by the guardrail
(llm_policy.model_blocked) before the upstream. The Vertex case mirrors
the customer verbatim (allow Sonnet; the unselected model is the exact
claude-opus-4-6 they reported reaching unblocked); the Bedrock case sends
a region-prefixed, versioned inference-profile id so URL-path model
normalization is exercised too.

The provider is catch-all, so the router forwards any model and a 403 can
only be the guardrail, never model_not_routable. Only the upstream LLM is
mocked (the vLLM nginx answers any path 200); management synth/reconcile,
the proxy middleware chain (URL-path model extraction, router, guardrail)
and the tunnel are all real, and the guardrail denies before the upstream
is dialed so the mock cannot influence the block. A static bearer api key
makes the router inject a static Authorization header instead of minting a
GCP token, so the test runs with no credentials and always executes.
This commit is contained in:
Maycon Santos
2026-07-23 05:01:40 +00:00
parent 178e6a8530
commit c48347c51e

View File

@@ -0,0 +1,209 @@
//go:build e2e
package agentnetwork
import (
"context"
"testing"
"time"
"github.com/stretchr/testify/assert"
"github.com/stretchr/testify/require"
"github.com/netbirdio/netbird/e2e/harness"
"github.com/netbirdio/netbird/shared/management/http/api"
)
// pathRoutedGuardrailCase is one provider's self-contained scenario: its own
// provider, its own guardrail whose allowlist holds ONLY that provider's
// allowed model, and its own policy. Each case runs in isolation (its own
// proxy + client), so the guardrail the proxy enforces contains exactly this
// provider's model — never a mixed cross-provider list.
type pathRoutedGuardrailCase struct {
name string
catalogID string // agent-network catalog provider id
wire string // harness.WireVertex | harness.WireBedrock
allowEntry string // the single model id put on the guardrail allowlist
allowModel string // model id sent that MUST be served (200)
blockModel string // model id sent that MUST be denied (403 model_blocked)
}
// TestGuardrailBlocksUnselectedModel_PathRouted is the end-to-end regression
// guard for the customer report that a model-allowlist guardrail attached to a
// policy has no effect for PATH-ROUTED providers — where the model travels in
// the URL, not the JSON body: Google Vertex (…/models/{model}:rawPredict) and
// AWS Bedrock (/model/{id}/invoke).
//
// Each provider is tested in isolation with a guardrail allowlisting a single
// model of its own: the allowed model (in the URL path) is served (200) and an
// unselected model (in the URL path) is denied 403 by the guardrail
// (llm_policy.model_blocked) before the upstream. The Vertex case mirrors the
// customer verbatim — allow Sonnet, and the unselected model is the exact
// claude-opus-4-6 they reported reaching the model unblocked. The Bedrock case
// sends a region-prefixed, versioned inference-profile id so URL-path model
// normalization is exercised too.
//
// The provider is catch-all (no models), so the router forwards any model and a
// 403 can only come from the guardrail, never model_not_routable. Only the
// upstream LLM is mocked (the vLLM nginx answers any path with 200); management
// synth/reconcile, the proxy middleware chain (URL-path model extraction,
// router, guardrail) and the tunnel are all real, and the guardrail denies
// before the upstream is dialed so the mock cannot influence the block. A
// static bearer api key is used so the router injects a static Authorization
// header instead of minting a GCP token — the only reason path-routed providers
// normally need live credentials — so the test runs with none and is always on.
func TestGuardrailBlocksUnselectedModel_PathRouted(t *testing.T) {
cases := []pathRoutedGuardrailCase{
{
name: "vertex",
catalogID: "vertex_ai_api",
wire: harness.WireVertex,
allowEntry: "claude-sonnet-4-5",
allowModel: "claude-sonnet-4-5",
blockModel: "claude-opus-4-6", // the customer-reported model
},
{
name: "bedrock",
catalogID: "bedrock_api",
wire: harness.WireBedrock,
allowEntry: "anthropic.claude-sonnet-4-5", // normalized catalog id
allowModel: "us.anthropic.claude-sonnet-4-5-v1:0",
blockModel: "us.anthropic.claude-opus-4-8-v1:0",
},
}
for _, tc := range cases {
tc := tc
t.Run(tc.name, func(t *testing.T) {
runPathRoutedGuardrailCase(t, tc)
})
}
}
func runPathRoutedGuardrailCase(t *testing.T, tc pathRoutedGuardrailCase) {
t.Helper()
const (
vertexProject = "e2e-project"
vertexRegion = "global"
)
ctx, cancel := context.WithTimeout(context.Background(), 15*time.Minute)
defer cancel()
vllm, err := harness.StartVLLM(ctx, srv)
require.NoError(t, err, "start mock upstream")
t.Cleanup(func() { _ = vllm.Terminate(context.Background()) })
grp, err := srv.API().Groups.Create(ctx, api.PostApiGroupsJSONRequestBody{Name: "e2e-guardrail-" + tc.name})
require.NoError(t, err, "create group")
t.Cleanup(func() { _ = srv.API().Groups.Delete(context.Background(), grp.Id) })
ephemeral := false
sk, err := srv.API().SetupKeys.Create(ctx, api.PostApiSetupKeysJSONRequestBody{
Name: "e2e-guardrail-" + tc.name + "-client",
Type: "reusable",
ExpiresIn: 86400,
UsageLimit: 0,
AutoGroups: []string{grp.Id},
Ephemeral: &ephemeral,
})
require.NoError(t, err, "mint setup key")
require.NotEmpty(t, sk.Key, "setup key plaintext")
// Catch-all provider (no models) so the router forwards any model; a static
// bearer key means the router injects a static auth header instead of minting
// a GCP token. Bootstraps the cluster if it isn't already.
staticKey := "static-e2e-token"
prov, err := srv.CreateProvider(ctx, api.AgentNetworkProviderRequest{
Name: tc.name,
ProviderId: tc.catalogID,
UpstreamUrl: vllm.URL,
ApiKey: &staticKey,
Enabled: ptr(true),
BootstrapCluster: ptr(harness.AgentNetworkCluster),
})
require.NoError(t, err, "create %s provider", tc.name)
t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), prov.Id) })
// Guardrail allowlisting ONLY this provider's allowed model.
var gr api.AgentNetworkGuardrailRequest
gr.Name = "e2e-guardrail-" + tc.name
gr.Checks.ModelAllowlist.Enabled = true
gr.Checks.ModelAllowlist.Models = []string{tc.allowEntry}
guard, err := srv.CreateGuardrail(ctx, gr)
require.NoError(t, err, "create guardrail")
t.Cleanup(func() { _ = srv.DeleteGuardrail(context.Background(), guard.Id) })
enabled := true
pol, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
Name: "e2e-guardrail-" + tc.name,
Enabled: &enabled,
SourceGroups: []string{grp.Id},
DestinationProviderIds: []string{prov.Id},
GuardrailIds: &[]string{guard.Id},
})
require.NoError(t, err, "create policy")
t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), pol.Id) })
settings, err := srv.GetSettings(ctx)
require.NoError(t, err, "read settings")
require.NotEmpty(t, settings.Endpoint, "endpoint must be assigned")
proxyToken, err := srv.CreateProxyTokenCLI(ctx, "e2e-guardrail-"+tc.name+"-proxy")
require.NoError(t, err, "mint proxy token")
px, err := harness.StartProxy(ctx, srv, proxyToken)
require.NoError(t, err, "start proxy")
t.Cleanup(func() { _ = px.Terminate(context.Background()) })
cl, err := harness.StartClient(ctx, srv, sk.Key)
require.NoError(t, err, "start client")
t.Cleanup(func() { _ = cl.Terminate(context.Background()) })
require.NoError(t, cl.WaitConnected(ctx, 90*time.Second), "client must connect to management")
// Probe first: the GET resolves the endpoint and its first packet wakes the
// lazy proxy peer, so WaitProxyPeer then observes it connected.
proxyIP, err := cl.ResolveProxyIP(ctx, settings.Endpoint)
require.NoError(t, err, "resolve endpoint to proxy IP")
if err := cl.WaitProxyPeer(ctx, 180*time.Second); err != nil {
t.Fatalf("client did not see the proxy peer: %v\n=== proxy logs ===\n%s", err, px.Logs(context.Background()))
}
send := func(model string) (int, string) {
var code int
var body string
var cerr error
switch tc.wire {
case harness.WireVertex:
code, body, cerr = cl.Vertex(ctx, settings.Endpoint, proxyIP, vertexProject, vertexRegion, model, "Reply with exactly: pong", "")
case harness.WireBedrock:
code, body, cerr = cl.Bedrock(ctx, settings.Endpoint, proxyIP, model, "Reply with exactly: pong", "")
default:
t.Fatalf("unsupported wire %q", tc.wire)
}
require.NoError(t, cerr, "request must reach the proxy for %s", tc.name)
return code, body
}
// Allowed model (in the URL path) is served. Retry to absorb tunnel/DNS
// jitter on the first call over the freshly warmed tunnel.
var code int
var body string
deadline := time.Now().Add(90 * time.Second)
for time.Now().Before(deadline) {
code, body = send(tc.allowModel)
if code == 200 {
break
}
time.Sleep(5 * time.Second)
}
assert.Equal(t, 200, code,
"allowed %s model (URL path) must be served; body: %s\n=== proxy logs ===\n%s", tc.name, body, px.Logs(context.Background()))
// Unselected model (in the URL path) must be blocked by the guardrail.
code, body = send(tc.blockModel)
assert.Equal(t, 403, code,
"unselected %s model (URL path) must be denied, not served; body: %s\n=== proxy logs ===\n%s", tc.name, body, px.Logs(context.Background()))
assert.Contains(t, body, "llm_policy.model_blocked",
"%s denial must come from the guardrail allowlist, not routing; body: %s", tc.name, body)
}