agentnetwork: scope model allowlist per policy/group and provider

The model-allowlist guardrail was merged into one account-wide union
and enforced flat on every request, independent of which policy,
group, or provider actually authorised it. With multiple policies —
especially a mix of guardrailed and un-guardrailed ones — this
produced two defects:

  - false-allow: a model allowlisted for one group/provider leaked to
    any caller, because the flat union ignored which policy authorised
    the request; and
  - false-deny: a policy with no guardrail (intended unrestricted) had
    its traffic blocked by an unrelated policy's allowlist.

Make enforcement policy/group-aware, mirroring how llm_limit_check
already resolves caps:

  - Management (SelectPolicyForRequest) is authoritative. It now uses
    the request model (already carried on the wire in
    CheckLLMPolicyLimitsRequest.model, previously ignored) to keep only
    the applicable policies whose guardrails permit the model; a policy
    with no allowlist-enabled guardrail is unrestricted. When policies
    govern the (provider, groups) but none permits the model it denies
    with llm_policy.model_blocked. No proto change — the field already
    existed and the proxy already sends it.

  - The proxy llm_guardrail becomes a per-provider fail-closed backstop.
    The synthesiser emits a per-provider allowlist only when every
    policy authorising that provider restricts models; a provider
    reachable by any un-guardrailed policy is omitted (unrestricted).
    The middleware keys the allowlist by the resolved provider id and
    keeps the unknown/undetermined-model fail-closed behaviour so
    path-routed providers stay enforced when management is unreachable.

Unit tests cover the selector gate (block, allow, cross-group no-leak,
un-guardrailed policy unrestricted, undetermined-model fail-closed) and
the per-provider synthesis; an e2e proves all three properties end to
end over the tunnel. Docs updated for the new config shape and the
two-layer enforcement.

docs/agent-networks/01-end-to-end-flows.md

@@ -109,7 +109,7 @@ sequenceDiagram
         Chk->>Inj: continue
         Inj->>Inj: inject NetBird identity headers per provider config
         Inj->>Grd: continue
-        Grd->>Grd: enforce model allowlist
+        Grd->>Grd: enforce per-provider allowlist (fail-closed backstop)
         Grd->>Up: forward (over WireGuard)
         Up-->>Resp: response (JSON or SSE stream)
         Resp->>Resp: parse usage tokens, completion
@@ -135,6 +135,15 @@ sequenceDiagram
   (`redact_pii = settings.RedactPii`). Phones, emails, credit cards,
   PII names — see `redact.go` for the full set. See
   [`modules/31-proxy-middleware-builtin.md`](modules/31-proxy-middleware-builtin.md).
+- The model allowlist is enforced in TWO places. `CheckLLMPolicyLimits`
+  is authoritative: it resolves the policy that governs this
+  (provider, caller-groups) and denies (`deny_code = llm_policy.model_blocked`)
+  when no applicable policy permits the model — so an allowlist scoped to
+  one group/provider never leaks to another, and an un-guardrailed policy
+  is genuinely unrestricted. `llm_guardrail` is a per-provider fail-closed
+  backstop: it only carries an allowlist for a provider every authorising
+  policy restricts, and blocks unknown/undetermined models even when
+  management is unreachable.
 - SSE streaming requires special handling on the response side; the
   parser must handle partial chunks without buffering the whole
   stream. See [`modules/32-proxy-llm-parsers.md`](modules/32-proxy-llm-parsers.md).

docs/agent-networks/modules/21-management-agentnetwork.md

@@ -122,7 +122,7 @@ At request time the path is independent: the proxy calls `SelectPolicyForRequest
   | on_request | 1 | `llm_router` | `{"providers":[{id, models[], upstream_*, auth_header_*, allowed_group_ids[]}]}` | **true** |
   | on_request | 2 | `llm_limit_check` | `{}` | – |
   | on_request | 3 | `llm_identity_inject` | `{"providers":[{provider_id, header_pair?, json_metadata?, extra_headers?}]}` | **true** |
-  | on_request | 4 | `llm_guardrail` | `{"model_allowlist"?, "prompt_capture":{enabled,redact_pii}}` | – |
+  | on_request | 4 | `llm_guardrail` | `{"provider_allowlists"?: {providerID: []model}, "prompt_capture":{enabled,redact_pii}}` | – |
   | on_response | 5 | `llm_limit_record` | `{}` (runs LAST at runtime) | – |
   | on_response | 6 | `cost_meter` | `{}` | – |
   | on_response | 7 | `llm_response_parser` | `{"capture_completion": <bool>, "redact_pii"?: true}` | – |

docs/agent-networks/modules/31-proxy-middleware-builtin.md

@@ -244,7 +244,7 @@ no mocks. Tests: `TestChain_AllowPath_StampsAttributionAndRecordsCounter`
 | `llm_router` | `{providers: [{id, models, upstream_scheme, upstream_host, upstream_path?, auth_header_name, auth_header_value, allowed_group_ids}]}` |
 | `llm_limit_check` | `{}` — pulls `MgmtClient` from `FactoryContext` |
 | `llm_identity_inject` | `{providers: [{provider_id, header_pair?|json_metadata?, extra_headers?}]}` |
-| `llm_guardrail` | `{model_allowlist: []string, prompt_capture: {enabled, redact_pii}}` |
+| `llm_guardrail` | `{provider_allowlists: {providerID: []string}, prompt_capture: {enabled, redact_pii}}` — allowlist keyed by resolved provider id; a provider absent from the map is unrestricted (fail-closed backstop; authoritative per-policy/group check is management's `CheckLLMPolicyLimits`) |
 | `llm_response_parser` | `{redact_pii?, capture_completion?: *bool}` |
 | `cost_meter` | `{pricing_path?}` (basename inside data-dir; defaults `pricing.yaml`) |
 | `llm_limit_record` | `{}` — same pattern as `llm_limit_check` |

e2e/agentnetwork/guardrail_multipolicy_test.go

@@ -0,0 +1,232 @@
+//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"
+)
+
+// TestGuardrailMultiPolicyModelAllowlist is the end-to-end regression guard for
+// the multi-policy interactions of the model-allowlist guardrail: what happens
+// when several enabled policies govern an account and some carry a guardrail
+// while others don't. The earlier implementation merged every policy's allowlist
+// into ONE account-wide union enforced flat on every request, which produced two
+// defects this test pins:
+//
+//   - false-ALLOW (cross-group leak): a model allowlisted only for another
+//     group's policy became usable by any caller, because the union ignored
+//     which policy/group actually authorised the request; and
+//   - false-DENY: a policy with NO guardrail (intended unrestricted) still had
+//     its traffic blocked by some other policy's allowlist.
+//
+// The fix scopes enforcement to the matched policy/group in management
+// (SelectPolicyForRequest) with a per-provider fail-closed backstop at the
+// proxy. The account here has one client in grpMain and three policies over two
+// catch-all upstreams (the mock vLLM answers any model 200, so a 403 can only be
+// a policy decision, never a routing miss):
+//
+//   - polMain : grpMain  -> pRestricted, guardrail allowlisting modelSelected
+//   - polOther: grpOther -> pRestricted, guardrail allowlisting modelOther
+//   - polOpen : grpMain  -> pOpen,       NO guardrail (unrestricted)
+//
+// Providers declare their models so routing is deterministic. Over the tunnel,
+// as the grpMain client:
+//
+//   - modelSelected on pRestricted is served (200) — allowed by grpMain's policy;
+//   - modelOther on pRestricted is denied 403 (llm_policy.model_blocked) — it is
+//     allowlisted only for grpOther and must NOT leak to grpMain; and
+//   - openModel on pOpen is served (200) — the un-guardrailed policy leaves that
+//     provider unrestricted and must NOT be blocked by another policy's list.
+//
+// Under the old account-wide union the middle case returned 200 (the leak) and
+// the last case returned 403 (the false-deny); both are inverted here.
+func TestGuardrailMultiPolicyModelAllowlist(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 15*time.Minute)
+	defer cancel()
+
+	const (
+		modelSelected = "e2e-selected"
+		modelOther    = "e2e-other"
+		openModel     = "e2e-open"
+	)
+
+	vllm, err := harness.StartVLLM(ctx, srv)
+	require.NoError(t, err, "start mock upstream")
+	t.Cleanup(func() { _ = vllm.Terminate(context.Background()) })
+
+	grpMain, err := srv.API().Groups.Create(ctx, api.PostApiGroupsJSONRequestBody{Name: "e2e-guardrail-mp-main"})
+	require.NoError(t, err, "create main group")
+	t.Cleanup(func() { _ = srv.API().Groups.Delete(context.Background(), grpMain.Id) })
+
+	grpOther, err := srv.API().Groups.Create(ctx, api.PostApiGroupsJSONRequestBody{Name: "e2e-guardrail-mp-other"})
+	require.NoError(t, err, "create other group")
+	t.Cleanup(func() { _ = srv.API().Groups.Delete(context.Background(), grpOther.Id) })
+
+	ephemeral := false
+	sk, err := srv.API().SetupKeys.Create(ctx, api.PostApiSetupKeysJSONRequestBody{
+		Name:       "e2e-guardrail-mp-client",
+		Type:       "reusable",
+		ExpiresIn:  86400,
+		UsageLimit: 0,
+		AutoGroups: []string{grpMain.Id}, // client joins grpMain only
+		Ephemeral:  &ephemeral,
+	})
+	require.NoError(t, err, "mint setup key")
+	require.NotEmpty(t, sk.Key, "setup key plaintext")
+
+	staticKey := "static-e2e-token"
+	models := func(ids ...string) *[]api.AgentNetworkProviderModel {
+		out := make([]api.AgentNetworkProviderModel, 0, len(ids))
+		for _, id := range ids {
+			out = append(out, api.AgentNetworkProviderModel{Id: id, InputPer1k: 0.001, OutputPer1k: 0.001})
+		}
+		return &out
+	}
+
+	// pRestricted declares the two guardrailed models so routing is deterministic
+	// (model -> provider). Created first, so it carries the bootstrap cluster.
+	pRestricted, err := srv.CreateProvider(ctx, api.AgentNetworkProviderRequest{
+		Name:             "restricted",
+		ProviderId:       "openai_api",
+		UpstreamUrl:      vllm.URL,
+		ApiKey:           &staticKey,
+		Enabled:          ptr(true),
+		Models:           models(modelSelected, modelOther),
+		BootstrapCluster: ptr(harness.AgentNetworkCluster),
+	})
+	require.NoError(t, err, "create restricted provider")
+	t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), pRestricted.Id) })
+
+	pOpen, err := srv.CreateProvider(ctx, api.AgentNetworkProviderRequest{
+		Name:        "open",
+		ProviderId:  "openai_api",
+		UpstreamUrl: vllm.URL,
+		ApiKey:      &staticKey,
+		Enabled:     ptr(true),
+		Models:      models(openModel),
+	})
+	require.NoError(t, err, "create open provider")
+	t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), pOpen.Id) })
+
+	mkGuardrail := func(name, model string) api.AgentNetworkGuardrail {
+		var gr api.AgentNetworkGuardrailRequest
+		gr.Name = name
+		gr.Checks.ModelAllowlist.Enabled = true
+		gr.Checks.ModelAllowlist.Models = []string{model}
+		g, gerr := srv.CreateGuardrail(ctx, gr)
+		require.NoError(t, gerr, "create guardrail %s", name)
+		t.Cleanup(func() { _ = srv.DeleteGuardrail(context.Background(), g.Id) })
+		return g
+	}
+	gMain := mkGuardrail("e2e-guardrail-mp-main", modelSelected)
+	gOther := mkGuardrail("e2e-guardrail-mp-other", modelOther)
+
+	enabled := true
+	// polMain: grpMain restricted to modelSelected on pRestricted.
+	polMain, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
+		Name:                   "e2e-guardrail-mp-main",
+		Enabled:                &enabled,
+		SourceGroups:           []string{grpMain.Id},
+		DestinationProviderIds: []string{pRestricted.Id},
+		GuardrailIds:           &[]string{gMain.Id},
+	})
+	require.NoError(t, err, "create main policy")
+	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), polMain.Id) })
+
+	// polOther: grpOther restricted to modelOther on the SAME provider. The
+	// client is not in grpOther, so modelOther must never be usable by it.
+	polOther, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
+		Name:                   "e2e-guardrail-mp-other",
+		Enabled:                &enabled,
+		SourceGroups:           []string{grpOther.Id},
+		DestinationProviderIds: []string{pRestricted.Id},
+		GuardrailIds:           &[]string{gOther.Id},
+	})
+	require.NoError(t, err, "create other policy")
+	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), polOther.Id) })
+
+	// polOpen: grpMain on pOpen with NO guardrail — unrestricted.
+	polOpen, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
+		Name:                   "e2e-guardrail-mp-open",
+		Enabled:                &enabled,
+		SourceGroups:           []string{grpMain.Id},
+		DestinationProviderIds: []string{pOpen.Id},
+	})
+	require.NoError(t, err, "create open policy")
+	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), polOpen.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-mp-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")
+	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) {
+		code, body, cerr := cl.Chat(ctx, settings.Endpoint, proxyIP, harness.WireChat, model, "Reply with exactly: pong", "")
+		require.NoError(t, cerr, "request must reach the proxy")
+		return code, body
+	}
+	// sendUntil200 absorbs first-call tunnel/DNS jitter on the freshly warmed tunnel.
+	sendUntil200 := func(model string) (int, string) {
+		var code int
+		var body string
+		deadline := time.Now().Add(90 * time.Second)
+		for time.Now().Before(deadline) {
+			code, body = send(model)
+			if code == 200 {
+				break
+			}
+			time.Sleep(5 * time.Second)
+		}
+		return code, body
+	}
+
+	t.Run("selected model allowed for its group", func(t *testing.T) {
+		code, body := sendUntil200(modelSelected)
+		assert.Equal(t, 200, code,
+			"grpMain's allowlisted model must be served; body: %s\n=== proxy logs ===\n%s", body, px.Logs(context.Background()))
+	})
+
+	t.Run("other group's model does not leak", func(t *testing.T) {
+		// modelOther is allowlisted only for grpOther. The grpMain client must be
+		// denied by management's per-policy/group check — not waved through by an
+		// account-wide union. This is the security-critical wrong-ALLOW guard.
+		code, body := send(modelOther)
+		assert.Equal(t, 403, code,
+			"another group's allowlisted model must be denied for this caller; body: %s\n=== proxy logs ===\n%s", body, px.Logs(context.Background()))
+		assert.Contains(t, body, "llm_policy.model_blocked",
+			"denial must be a model-allowlist decision; body: %s", body)
+	})
+
+	t.Run("unguarded policy leaves its provider unrestricted", func(t *testing.T) {
+		// polOpen carries no guardrail, so pOpen is unrestricted for grpMain. The
+		// old account-wide union would have blocked openModel (it is on no
+		// allowlist); it must now be served — the false-DENY guard.
+		code, body := sendUntil200(openModel)
+		assert.Equal(t, 200, code,
+			"an un-guardrailed policy's provider must not be blocked by another policy's allowlist; body: %s\n=== proxy logs ===\n%s", body, px.Logs(context.Background()))
+	})
+}

management/internals/modules/agentnetwork/manager.go

@@ -86,12 +86,20 @@ type Manager interface {
 
 // PolicySelectionInput is the per-request selection envelope. The
 // proxy populates it from CapturedData (account, user, groups) plus
-// the provider llm_router resolved.
+// the provider llm_router resolved and the model it extracted.
 type PolicySelectionInput struct {
 	AccountID  string
 	UserID     string
 	GroupIDs   []string
 	ProviderID string
+	// Model is the (already-normalised) upstream model identifier the proxy
+	// extracted from the request. The proxy's request parser strips
+	// path-routed provider decoration (Bedrock region/version, Vertex
+	// @version) before emitting it, so a plain case-insensitive compare
+	// against a guardrail allowlist is sufficient here. Empty when the model
+	// could not be determined — treated as "not permitted" by any policy that
+	// restricts models (fail closed), mirroring the proxy guardrail.
+	Model string
 }
 
 // PolicySelectionResult names the policy that "pays" for this request

management/internals/modules/agentnetwork/policyselect.go

@@ -5,6 +5,7 @@ import (
 	"fmt"
 	"math"
 	"sort"
+	"strings"
 	"time"
 
 	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
@@ -35,6 +36,11 @@ const (
 	denyCodeAccountTokenCapExceeded = "llm_account.token_cap_exceeded"
 	//nolint:gosec // account deny code label, not a credential
 	denyCodeAccountBudgetCapExceeded = "llm_account.budget_cap_exceeded"
+	// denyCodeModelBlocked is returned when policies govern the request's
+	// (provider, caller-groups) but none of them permits the requested model
+	// under its guardrail allowlist. Matches the proxy guardrail's code so the
+	// two enforcement layers surface the same label.
+	denyCodeModelBlocked = "llm_policy.model_blocked"
 )
 
 // consumptionCache holds the consumption counters prefetched for one
@@ -159,6 +165,34 @@ func (m *managerImpl) SelectPolicyForRequest(ctx context.Context, in PolicySelec
 	}
 	candidates := filterApplicablePolicies(policies, in)
 
+	// Model-allowlist gate (per-policy, per-group). Among the policies that
+	// authorise this (provider, caller-groups), keep only those whose guardrails
+	// permit the requested model; a policy with no allowlist-enabled guardrail is
+	// unrestricted and always permits. When policies govern this request but none
+	// permits the model, deny. This is the authoritative allowlist decision:
+	// because it is scoped to the matched policies, a model allowlisted for one
+	// group/provider never leaks to another (no account-wide union), and a policy
+	// that carries no guardrail is genuinely unrestricted rather than being caught
+	// by some other policy's allowlist.
+	// Only consult guardrails when at least one candidate policy references one;
+	// a policy with no guardrail is unrestricted, so when none of them carries a
+	// guardrail the gate is a no-op and we skip the store read entirely.
+	if len(candidates) > 0 && anyPolicyHasGuardrails(candidates) {
+		guardrailsByID, gErr := m.loadGuardrailsByID(ctx, in.AccountID)
+		if gErr != nil {
+			return nil, gErr
+		}
+		permitted := filterModelPermittedPolicies(candidates, guardrailsByID, in.Model)
+		if len(permitted) == 0 {
+			return &PolicySelectionResult{
+				Allow:      false,
+				DenyCode:   denyCodeModelBlocked,
+				DenyReason: modelBlockedReason(in.Model),
+			}, nil
+		}
+		candidates = permitted
+	}
+
 	// Prefetch every consumption counter the ceiling + candidate policies will
 	// read, in a single store round-trip, then score against the cache.
 	cache, err := m.prefetchConsumption(ctx, in, rules, candidates, now)
@@ -250,6 +284,93 @@ func filterApplicablePolicies(policies []*types.Policy, in PolicySelectionInput)
 	return out
 }
 
+// anyPolicyHasGuardrails reports whether any policy references at least one
+// guardrail, so the selector can skip loading guardrails when none do.
+func anyPolicyHasGuardrails(policies []*types.Policy) bool {
+	for _, p := range policies {
+		if p != nil && len(p.GuardrailIDs) > 0 {
+			return true
+		}
+	}
+	return false
+}
+
+// loadGuardrailsByID loads the account's guardrails indexed by ID. Used by the
+// model-allowlist gate to resolve each candidate policy's attached guardrails.
+func (m *managerImpl) loadGuardrailsByID(ctx context.Context, accountID string) (map[string]*types.Guardrail, error) {
+	guardrails, err := m.store.GetAccountAgentNetworkGuardrails(ctx, store.LockingStrengthNone, accountID)
+	if err != nil {
+		return nil, fmt.Errorf("list account guardrails: %w", err)
+	}
+	byID := make(map[string]*types.Guardrail, len(guardrails))
+	for _, g := range guardrails {
+		if g != nil {
+			byID[g.ID] = g
+		}
+	}
+	return byID, nil
+}
+
+// filterModelPermittedPolicies returns the subset of policies whose guardrails
+// permit the model. Order is preserved so downstream scoring is unaffected.
+func filterModelPermittedPolicies(policies []*types.Policy, byID map[string]*types.Guardrail, model string) []*types.Policy {
+	out := make([]*types.Policy, 0, len(policies))
+	for _, p := range policies {
+		if policyPermitsModel(p, byID, model) {
+			out = append(out, p)
+		}
+	}
+	return out
+}
+
+// policyPermitsModel reports whether a policy permits the requested model under
+// its attached guardrails. A policy with no allowlist-enabled guardrail is
+// unrestricted (permits any model, including an empty/undetermined one). A
+// policy with one or more allowlist-enabled guardrails permits the model only
+// when it appears in the union of those allowlists; an empty/undetermined model
+// never matches a non-empty allowlist, so such a policy fails closed — the same
+// contract the proxy guardrail enforces for path-routed providers.
+func policyPermitsModel(p *types.Policy, byID map[string]*types.Guardrail, model string) bool {
+	if p == nil {
+		return false
+	}
+	wanted := normaliseModelID(model)
+	restricted := false
+	for _, gID := range p.GuardrailIDs {
+		g, ok := byID[gID]
+		if !ok || g == nil || !g.Checks.ModelAllowlist.Enabled {
+			continue
+		}
+		restricted = true
+		if wanted == "" {
+			continue
+		}
+		for _, allowed := range g.Checks.ModelAllowlist.Models {
+			if normaliseModelID(allowed) == wanted {
+				return true
+			}
+		}
+	}
+	return !restricted
+}
+
+// normaliseModelID lowercases and trims a model identifier so the allowlist
+// compare is case-insensitive and trim-tolerant. Mirrors the proxy guardrail's
+// normaliseModel so both layers agree on what "same model" means.
+func normaliseModelID(model string) string {
+	return strings.ToLower(strings.TrimSpace(model))
+}
+
+// modelBlockedReason builds the human-readable deny reason for a model-allowlist
+// rejection. The model is quoted when known; an undetermined model is reported
+// as such so the access log distinguishes "wrong model" from "no model".
+func modelBlockedReason(model string) string {
+	if normaliseModelID(model) == "" {
+		return "request model could not be determined for the policy allowlist"
+	}
+	return fmt.Sprintf("model %q is not permitted by any applicable policy allowlist", model)
+}
+
 // candidate is the per-policy intermediate the selector ranks. A
 // policy that's been exhausted on any enabled cap never makes it
 // into this slice; the selector's deny envelope carries the latest

management/internals/modules/agentnetwork/policyselect_model_test.go

@@ -0,0 +1,226 @@
+package agentnetwork
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"github.com/golang/mock/gomock"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
+	"github.com/netbirdio/netbird/management/server/store"
+)
+
+// guardedPolicy builds an enabled, uncapped policy that authorises sourceGroups
+// to reach providerID under the given guardrails. Uncapped keeps the selector's
+// headroom scoring trivial so these tests isolate the model-allowlist gate.
+func guardedPolicy(id, account string, sourceGroups []string, providerID string, guardrailIDs ...string) *types.Policy {
+	return &types.Policy{
+		ID:                     id,
+		AccountID:              account,
+		Enabled:                true,
+		SourceGroups:           sourceGroups,
+		DestinationProviderIDs: []string{providerID},
+		GuardrailIDs:           guardrailIDs,
+		CreatedAt:              time.Now().UTC(),
+	}
+}
+
+// allowlistGuardrail builds a guardrail whose model allowlist is enabled and
+// carries the given models.
+func allowlistGuardrail(id, account string, models ...string) *types.Guardrail {
+	return &types.Guardrail{
+		ID:        id,
+		AccountID: account,
+		Checks: types.GuardrailChecks{
+			ModelAllowlist: types.GuardrailModelAllowlist{Enabled: true, Models: models},
+		},
+	}
+}
+
+func expectPolicies(mockStore *store.MockStore, account string, policies ...*types.Policy) {
+	mockStore.EXPECT().
+		GetAccountAgentNetworkPolicies(gomock.Any(), gomock.Any(), account).
+		Return(policies, nil)
+}
+
+func expectGuardrails(mockStore *store.MockStore, account string, guardrails ...*types.Guardrail) {
+	mockStore.EXPECT().
+		GetAccountAgentNetworkGuardrails(gomock.Any(), gomock.Any(), account).
+		Return(guardrails, nil)
+}
+
+// TestSelectPolicy_ModelBlockedByAllowlist proves the authoritative allowlist
+// decision: a policy authorises the (provider, group) but restricts the model,
+// and the requested model isn't on the list, so the request is denied.
+func TestSelectPolicy_ModelBlockedByAllowlist(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-1")
+	expectPolicies(mockStore, "acc-1", policy)
+	expectGuardrails(mockStore, "acc-1", allowlistGuardrail("g-1", "acc-1", "gpt-4o"))
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		UserID:     "user-1",
+		GroupIDs:   []string{"grp-eng"},
+		ProviderID: "prov-1",
+		Model:      "claude-opus-4",
+	})
+	require.NoError(t, err)
+	assert.False(t, res.Allow, "a model outside the only applicable policy's allowlist must be denied")
+	assert.Equal(t, denyCodeModelBlocked, res.DenyCode, "deny code must be model_blocked")
+	assert.NotEmpty(t, res.DenyReason, "deny reason must be populated")
+}
+
+// TestSelectPolicy_ModelAllowedByAllowlist is the allow counterpart: the model
+// is on the applicable policy's allowlist, so selection proceeds normally.
+func TestSelectPolicy_ModelAllowedByAllowlist(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-1")
+	expectPolicies(mockStore, "acc-1", policy)
+	expectGuardrails(mockStore, "acc-1", allowlistGuardrail("g-1", "acc-1", "gpt-4o", "claude-opus-4"))
+	expectConsumptionBatch(mockStore, nil)
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		UserID:     "user-1",
+		GroupIDs:   []string{"grp-eng"},
+		ProviderID: "prov-1",
+		Model:      "claude-opus-4",
+	})
+	require.NoError(t, err)
+	assert.True(t, res.Allow, "a model on the applicable policy's allowlist must be allowed")
+	assert.Equal(t, "pol-A", res.SelectedPolicyID)
+}
+
+// TestSelectPolicy_CaseInsensitiveModelMatch proves the compare tolerates case
+// and surrounding whitespace, matching the proxy guardrail's normalisation.
+func TestSelectPolicy_CaseInsensitiveModelMatch(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-1")
+	expectPolicies(mockStore, "acc-1", policy)
+	expectGuardrails(mockStore, "acc-1", allowlistGuardrail("g-1", "acc-1", "  GPT-4o  "))
+	expectConsumptionBatch(mockStore, nil)
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		GroupIDs:   []string{"grp-eng"},
+		ProviderID: "prov-1",
+		Model:      "gpt-4o",
+	})
+	require.NoError(t, err)
+	assert.True(t, res.Allow, "case/whitespace variants must match the allowlist entry")
+}
+
+// TestSelectPolicy_UnguardedPolicyIsUnrestricted is the false-deny fix: when
+// two policies authorise the same (provider, group) and one carries no
+// guardrail, that un-guardrailed policy makes the request unrestricted — it is
+// NOT caught by the other policy's allowlist.
+func TestSelectPolicy_UnguardedPolicyIsUnrestricted(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	restricted := guardedPolicy("pol-restricted", "acc-1", []string{"grp-eng"}, "prov-1", "g-1")
+	open := guardedPolicy("pol-open", "acc-1", []string{"grp-eng"}, "prov-1") // no guardrail
+	expectPolicies(mockStore, "acc-1", restricted, open)
+	expectGuardrails(mockStore, "acc-1", allowlistGuardrail("g-1", "acc-1", "gpt-4o"))
+	expectConsumptionBatch(mockStore, nil)
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		GroupIDs:   []string{"grp-eng"},
+		ProviderID: "prov-1",
+		Model:      "claude-opus-4",
+	})
+	require.NoError(t, err)
+	assert.True(t, res.Allow, "an un-guardrailed policy for the same (provider, group) must leave the request unrestricted")
+	assert.Equal(t, "pol-open", res.SelectedPolicyID, "the unrestricted policy must be the one that pays")
+}
+
+// TestSelectPolicy_AllowlistDoesNotLeakAcrossGroups is the false-allow fix: a
+// model allowlisted only for grp-b must not be usable by a caller in grp-a,
+// even though both groups' policies target the same provider. The selector only
+// considers policies applicable to the caller's groups, so grp-b's allowlist
+// never enters grp-a's decision.
+func TestSelectPolicy_AllowlistDoesNotLeakAcrossGroups(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	polA := guardedPolicy("pol-a", "acc-1", []string{"grp-a"}, "prov-1", "g-a")
+	polB := guardedPolicy("pol-b", "acc-1", []string{"grp-b"}, "prov-1", "g-b")
+	expectPolicies(mockStore, "acc-1", polA, polB)
+	expectGuardrails(mockStore, "acc-1",
+		allowlistGuardrail("g-a", "acc-1", "gpt-4o"),
+		allowlistGuardrail("g-b", "acc-1", "claude-opus-4"),
+	)
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		GroupIDs:   []string{"grp-a"},
+		ProviderID: "prov-1",
+		Model:      "claude-opus-4", // only allowed for grp-b
+	})
+	require.NoError(t, err)
+	assert.False(t, res.Allow, "grp-b's allowlisted model must not leak to a grp-a caller")
+	assert.Equal(t, denyCodeModelBlocked, res.DenyCode)
+}
+
+// TestSelectPolicy_UndeterminedModelFailsClosed proves the fail-closed contract
+// mirrors the proxy: with a restricted applicable policy and an empty model
+// (e.g. a path-routed shape the parser couldn't map), the request is denied.
+func TestSelectPolicy_UndeterminedModelFailsClosed(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-1")
+	expectPolicies(mockStore, "acc-1", policy)
+	expectGuardrails(mockStore, "acc-1", allowlistGuardrail("g-1", "acc-1", "gpt-4o"))
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		GroupIDs:   []string{"grp-eng"},
+		ProviderID: "prov-1",
+		Model:      "", // undetermined
+	})
+	require.NoError(t, err)
+	assert.False(t, res.Allow, "an undetermined model must fail closed against a restricted policy")
+	assert.Equal(t, denyCodeModelBlocked, res.DenyCode)
+}
+
+// TestSelectPolicy_DisabledAllowlistDoesNotRestrict proves a guardrail whose
+// model allowlist is disabled imposes no model restriction, even though the
+// policy references it.
+func TestSelectPolicy_DisabledAllowlistDoesNotRestrict(t *testing.T) {
+	ctrl := gomock.NewController(t)
+	mgr, mockStore := newSelectorMgr(t, ctrl)
+
+	policy := guardedPolicy("pol-A", "acc-1", []string{"grp-eng"}, "prov-1", "g-1")
+	disabled := &types.Guardrail{
+		ID:        "g-1",
+		AccountID: "acc-1",
+		Checks: types.GuardrailChecks{
+			ModelAllowlist: types.GuardrailModelAllowlist{Enabled: false, Models: []string{"gpt-4o"}},
+		},
+	}
+	expectPolicies(mockStore, "acc-1", policy)
+	expectGuardrails(mockStore, "acc-1", disabled)
+	expectConsumptionBatch(mockStore, nil)
+
+	res, err := mgr.SelectPolicyForRequest(context.Background(), PolicySelectionInput{
+		AccountID:  "acc-1",
+		GroupIDs:   []string{"grp-eng"},
+		ProviderID: "prov-1",
+		Model:      "anything-goes",
+	})
+	require.NoError(t, err)
+	assert.True(t, res.Allow, "a disabled allowlist must not restrict the model")
+	assert.Equal(t, "pol-A", res.SelectedPolicyID)
+}

management/internals/modules/agentnetwork/synthesizer.go

@@ -235,7 +235,14 @@ func SynthesizeServices(ctx context.Context, s store.Store, accountID string) ([
 
 	mergedGuardrails := mergeGuardrails(enabledPolicies, guardrailsByID)
 	applyAccountCollectionControls(&mergedGuardrails, settings)
-	guardrailJSON, err := marshalGuardrailConfig(mergedGuardrails)
+	// The proxy guardrail is the fail-closed backstop, scoped per provider so it
+	// never applies one policy's allowlist to another provider's traffic. The
+	// authoritative per-policy/group model decision lives in management's
+	// SelectPolicyForRequest; a provider only lands in this map when EVERY policy
+	// authorising it restricts models, so an un-guardrailed policy leaves its
+	// providers unrestricted here.
+	providerAllowlists := buildProviderAllowlists(enabledPolicies, guardrailsByID)
+	guardrailJSON, err := marshalGuardrailConfig(providerAllowlists, mergedGuardrails.PromptCapture)
 	if err != nil {
 		return nil, err
 	}
@@ -780,10 +787,12 @@ func buildMiddlewareChain(routerCfgJSON, identityInjectJSON, guardrailJSON []byt
 
 // guardrailConfig is the JSON shape the proxy-side llm_guardrail
 // middleware expects. Mirrors the proxy registration documented in
-// the management→proxy contract.
+// the management→proxy contract. provider_allowlists is keyed by the
+// resolved provider id llm_router stamps; a provider absent from the map is
+// unrestricted at the proxy layer.
 type guardrailConfig struct {
-	ModelAllowlist []string               `json:"model_allowlist,omitempty"`
-	PromptCapture  guardrailPromptCapture `json:"prompt_capture"`
+	ProviderAllowlists map[string][]string    `json:"provider_allowlists,omitempty"`
+	PromptCapture      guardrailPromptCapture `json:"prompt_capture"`
 }
 
 type guardrailPromptCapture struct {
@@ -828,12 +837,12 @@ func applyAccountCollectionControls(merged *MergedGuardrails, settings *types.Se
 	merged.PromptCapture.RedactPii = settings.RedactPii || merged.PromptCapture.RedactPii
 }
 
-func marshalGuardrailConfig(merged MergedGuardrails) ([]byte, error) {
+func marshalGuardrailConfig(providerAllowlists map[string][]string, capture MergedPromptCapture) ([]byte, error) {
 	cfg := guardrailConfig{
-		ModelAllowlist: merged.ModelAllowlist,
+		ProviderAllowlists: providerAllowlists,
 		PromptCapture: guardrailPromptCapture{
-			Enabled:   merged.PromptCapture.Enabled,
-			RedactPii: merged.PromptCapture.RedactPii,
+			Enabled:   capture.Enabled,
+			RedactPii: capture.RedactPii,
 		},
 	}
 	out, err := json.Marshal(cfg)
@@ -843,6 +852,81 @@ func marshalGuardrailConfig(merged MergedGuardrails) ([]byte, error) {
 	return out, nil
 }
 
+// buildProviderAllowlists computes the per-provider model allowlist the proxy
+// guardrail enforces as a fail-closed backstop. A provider is included ONLY when
+// every enabled policy that authorises it restricts models (has at least one
+// allowlist-enabled guardrail); its list is then the union of those policies'
+// allowed models. If any authorising policy leaves models unrestricted, the
+// provider is omitted entirely — the proxy treats an absent provider as
+// unrestricted and defers to management's per-policy/group decision, so a
+// mixed set of policies can neither leak a model across providers nor wrongly
+// block an un-guardrailed policy's traffic.
+func buildProviderAllowlists(policies []*types.Policy, byID map[string]*types.Guardrail) map[string][]string {
+	type providerAcc struct {
+		models          map[string]struct{}
+		anyUnrestricted bool
+	}
+	accs := make(map[string]*providerAcc)
+	for _, p := range policies {
+		if p == nil {
+			continue
+		}
+		restricted, models := policyModelAllowlist(p, byID)
+		for _, providerID := range p.DestinationProviderIDs {
+			if providerID == "" {
+				continue
+			}
+			acc, ok := accs[providerID]
+			if !ok {
+				acc = &providerAcc{models: make(map[string]struct{})}
+				accs[providerID] = acc
+			}
+			if !restricted {
+				acc.anyUnrestricted = true
+				continue
+			}
+			for _, m := range models {
+				acc.models[m] = struct{}{}
+			}
+		}
+	}
+	out := make(map[string][]string, len(accs))
+	for providerID, acc := range accs {
+		if acc.anyUnrestricted {
+			continue
+		}
+		models := make([]string, 0, len(acc.models))
+		for m := range acc.models {
+			models = append(models, m)
+		}
+		sort.Strings(models)
+		out[providerID] = models
+	}
+	return out
+}
+
+// policyModelAllowlist reports whether a policy restricts models (has at least
+// one allowlist-enabled guardrail) and the union of allowed models across those
+// guardrails. Models are returned verbatim; the proxy factory lowercases and
+// trims them at decode time, matching the runtime compare.
+func policyModelAllowlist(p *types.Policy, byID map[string]*types.Guardrail) (bool, []string) {
+	restricted := false
+	var models []string
+	for _, gID := range p.GuardrailIDs {
+		g, ok := byID[gID]
+		if !ok || g == nil || !g.Checks.ModelAllowlist.Enabled {
+			continue
+		}
+		restricted = true
+		for _, m := range g.Checks.ModelAllowlist.Models {
+			if m != "" {
+				models = append(models, m)
+			}
+		}
+	}
+	return restricted, models
+}
+
 // buildAccountService composes the per-account gateway Service. The
 // target carries the noop placeholder URL — the router middleware
 // rewrites every request to the matched provider's upstream before the
@@ -991,11 +1075,10 @@ func unionSourceGroups(policies []*types.Policy) []string {
 // expectations and is intentionally distinct from
 // types.GuardrailChecks so we can evolve either side independently.
 type MergedGuardrails struct {
-	ModelAllowlist []string            `json:"model_allowlist,omitempty"`
-	TokenLimits    MergedTokenLimits   `json:"token_limits"`
-	Budget         MergedBudget        `json:"budget"`
-	PromptCapture  MergedPromptCapture `json:"prompt_capture"`
-	Retention      MergedRetention     `json:"retention"`
+	TokenLimits   MergedTokenLimits   `json:"token_limits"`
+	Budget        MergedBudget        `json:"budget"`
+	PromptCapture MergedPromptCapture `json:"prompt_capture"`
+	Retention     MergedRetention     `json:"retention"`
 }
 
 type MergedTokenLimits struct {
@@ -1030,59 +1113,38 @@ type MergedRetention struct {
 	Days    int  `json:"days"`
 }
 
-// mergeGuardrails computes the effective guardrail spec applied at the
-// proxy, given the referencing policies and the account's guardrail
-// catalogue. Policy enabled-ness is the caller's responsibility — only
-// enabled policies should be passed in.
+// mergeGuardrails computes the prompt-capture portion of the effective
+// guardrail spec, given the referencing policies and the account's guardrail
+// catalogue. Policy enabled-ness is the caller's responsibility — only enabled
+// policies should be passed in.
 //
-// Merge rules:
-//   - Model allowlist:   union of allowlists across policies that enable it.
-//   - Token / Budget:    most-restrictive (min of non-zero caps) per window.
-//   - Prompt capture:    enabled if any policy enables it; redact_pii sticks
-//     if any enabling policy turns it on.
-//   - Retention:         enabled if any enables it; smallest non-zero days wins.
+// The model allowlist is NOT merged here: it is enforced per-policy/group in
+// management (SelectPolicyForRequest) and shipped to the proxy per-provider via
+// buildProviderAllowlists, so a single account-wide union would be both
+// incorrect (leaks a model across groups/providers) and redundant. Token,
+// budget, and retention likewise live off guardrails now (Policy.Limits and
+// account Settings), so prompt capture is all that remains to merge.
+//
+// Merge rule — prompt capture: enabled if any policy enables it; redact_pii
+// sticks if any enabling policy turns it on.
 func mergeGuardrails(policies []*types.Policy, byID map[string]*types.Guardrail) MergedGuardrails {
 	merged := MergedGuardrails{}
-	allowlist := make(map[string]struct{})
-	allowlistEnabled := false
-
 	for _, policy := range policies {
 		for _, gID := range policy.GuardrailIDs {
 			g, ok := byID[gID]
 			if !ok || g == nil {
 				continue
 			}
-			mergeGuardrail(g, &merged, allowlist, &allowlistEnabled)
+			mergeGuardrail(g, &merged)
 		}
 	}
-
-	if allowlistEnabled {
-		merged.ModelAllowlist = make([]string, 0, len(allowlist))
-		for m := range allowlist {
-			merged.ModelAllowlist = append(merged.ModelAllowlist, m)
-		}
-		sort.Strings(merged.ModelAllowlist)
-	}
 	return merged
 }
 
-// mergeGuardrail folds a single guardrail's enabled checks into the
-// running merge: model-allowlist models join the shared set (and flip
-// allowlistEnabled), and prompt-capture / redact-pii stick once any
-// enabling guardrail turns them on.
-//
-// TokenLimits, Budget, and Retention have moved off guardrails — token
-// and budget caps now live on the Policy itself (Policy.Limits) and
-// retention moves to account-level Settings — so they are not merged here.
-func mergeGuardrail(g *types.Guardrail, merged *MergedGuardrails, allowlist map[string]struct{}, allowlistEnabled *bool) {
-	if g.Checks.ModelAllowlist.Enabled {
-		*allowlistEnabled = true
-		for _, m := range g.Checks.ModelAllowlist.Models {
-			if m != "" {
-				allowlist[m] = struct{}{}
-			}
-		}
-	}
+// mergeGuardrail folds a single guardrail's prompt-capture settings into the
+// running merge: enabled / redact-pii stick once any enabling guardrail turns
+// them on.
+func mergeGuardrail(g *types.Guardrail, merged *MergedGuardrails) {
 	if g.Checks.PromptCapture.Enabled {
 		merged.PromptCapture.Enabled = true
 		if g.Checks.PromptCapture.RedactPii {

management/internals/modules/agentnetwork/synthesizer_guardrail_realstore_test.go

@@ -81,8 +81,8 @@ func TestSynthesizeServices_RealStore_PromptCaptureAccountIsSoleControl(t *testi
 	require.Len(t, services, 1)
 
 	cfg := decodeServiceGuardrailConfig(t, services[0])
-	assert.Equal(t, []string{"gpt-5.4"}, cfg.ModelAllowlist,
-		"model allowlist is a pure policy guardrail and must always reach the config")
+	assert.Equal(t, map[string][]string{"prov-1": {"gpt-5.4"}}, cfg.ProviderAllowlists,
+		"model allowlist is a pure policy guardrail and must reach the per-provider config")
 	assert.False(t, cfg.PromptCapture.Enabled,
 		"prompt capture must be off when the account toggle is off, even with a capture-enabled guardrail")
 }
@@ -172,7 +172,7 @@ func TestSynthesizeServices_RealStore_NoGuardrail_CaptureOff(t *testing.T) {
 	require.Len(t, services, 1, "exactly one synth service expected")
 
 	cfg := decodeServiceGuardrailConfig(t, services[0])
-	assert.Empty(t, cfg.ModelAllowlist, "no guardrail → no allowlist")
+	assert.Empty(t, cfg.ProviderAllowlists, "no guardrail → provider unrestricted (absent from map)")
 	assert.False(t, cfg.PromptCapture.Enabled, "no guardrail → prompt capture off by default")
 	assert.False(t, cfg.PromptCapture.RedactPii, "no guardrail → redact off by default")
 }

management/internals/modules/agentnetwork/synthesizer_provider_allowlist_test.go

@@ -0,0 +1,77 @@
+package agentnetwork
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+
+	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
+)
+
+// policyForProviders builds an enabled policy authorising the given providers
+// under the given guardrails (both optional). Groups are irrelevant to
+// buildProviderAllowlists, which keys purely on destination provider.
+func policyForProviders(id string, guardrailIDs []string, providerIDs ...string) *types.Policy {
+	return &types.Policy{
+		ID:                     id,
+		Enabled:                true,
+		DestinationProviderIDs: providerIDs,
+		GuardrailIDs:           guardrailIDs,
+	}
+}
+
+func TestBuildProviderAllowlists(t *testing.T) {
+	byID := map[string]*types.Guardrail{
+		"g-4o":       allowlistGuardrail("g-4o", "acc-1", "gpt-4o"),
+		"g-opus":     allowlistGuardrail("g-opus", "acc-1", "claude-opus-4"),
+		"g-disabled": {ID: "g-disabled", Checks: types.GuardrailChecks{ModelAllowlist: types.GuardrailModelAllowlist{Enabled: false, Models: []string{"gpt-4o"}}}},
+	}
+
+	t.Run("all authorising policies restrict yields per-provider union", func(t *testing.T) {
+		policies := []*types.Policy{
+			policyForProviders("p1", []string{"g-4o"}, "prov-x"),
+			policyForProviders("p2", []string{"g-opus"}, "prov-x"),
+		}
+		got := buildProviderAllowlists(policies, byID)
+		assert.Equal(t, map[string][]string{"prov-x": {"claude-opus-4", "gpt-4o"}}, got,
+			"a provider every policy restricts carries the sorted union of their models")
+	})
+
+	t.Run("any un-guardrailed policy leaves the provider unrestricted (omitted)", func(t *testing.T) {
+		policies := []*types.Policy{
+			policyForProviders("p1", []string{"g-4o"}, "prov-x"),
+			policyForProviders("p2", nil, "prov-x"), // no guardrail
+		}
+		got := buildProviderAllowlists(policies, byID)
+		assert.NotContains(t, got, "prov-x",
+			"a provider reachable by an un-guardrailed policy must be omitted so the proxy treats it as unrestricted")
+	})
+
+	t.Run("a disabled allowlist counts as unrestricted", func(t *testing.T) {
+		policies := []*types.Policy{
+			policyForProviders("p1", []string{"g-disabled"}, "prov-x"),
+		}
+		got := buildProviderAllowlists(policies, byID)
+		assert.NotContains(t, got, "prov-x",
+			"a policy whose only guardrail has a disabled allowlist is unrestricted")
+	})
+
+	t.Run("providers are isolated from one another", func(t *testing.T) {
+		policies := []*types.Policy{
+			policyForProviders("p1", []string{"g-4o"}, "prov-x"),
+			policyForProviders("p2", []string{"g-opus"}, "prov-y"),
+		}
+		got := buildProviderAllowlists(policies, byID)
+		assert.Equal(t, []string{"gpt-4o"}, got["prov-x"], "prov-x keeps only its own model")
+		assert.Equal(t, []string{"claude-opus-4"}, got["prov-y"], "prov-y keeps only its own model")
+	})
+
+	t.Run("one policy authorising two providers restricts both", func(t *testing.T) {
+		policies := []*types.Policy{
+			policyForProviders("p1", []string{"g-4o"}, "prov-x", "prov-y"),
+		}
+		got := buildProviderAllowlists(policies, byID)
+		assert.Equal(t, []string{"gpt-4o"}, got["prov-x"])
+		assert.Equal(t, []string{"gpt-4o"}, got["prov-y"])
+	})
+}

management/internals/modules/agentnetwork/synthesizer_test.go

@@ -1031,8 +1031,14 @@ func TestSynthesizeServices_GuardrailMerge_AllowlistUnion_LimitsRestrictive(t *t
 
 	var cfg guardrailConfig
 	require.NoError(t, json.Unmarshal(guardrailJSON, &cfg), "guardrail config must unmarshal cleanly")
-	assert.ElementsMatch(t, []string{"gpt-5.4-mini", "gpt-5.4-pro"}, cfg.ModelAllowlist,
-		"model allowlist union must keep both models")
+	// Both policies restrict the same provider, so the proxy-side per-provider
+	// backstop carries the union of their models. This coarse union is
+	// deliberate: it can still let grp-a reach grp-b's model at the proxy layer,
+	// which management's per-policy/group check (SelectPolicyForRequest) is the
+	// one that rejects. The backstop only guarantees nothing outside the
+	// provider's union slips through when management is unreachable.
+	assert.ElementsMatch(t, []string{"gpt-5.4-mini", "gpt-5.4-pro"}, cfg.ProviderAllowlists["prov-1"],
+		"per-provider allowlist union must keep both models")
 }
 
 func TestSynthesizeServices_BackfillsMissingSessionKeys(t *testing.T) {

management/internals/shared/grpc/proxy.go

@@ -285,6 +285,7 @@ func (s *ProxyServiceServer) CheckLLMPolicyLimits(ctx context.Context, req *prot
 		UserID:     req.GetUserId(),
 		GroupIDs:   req.GetGroupIds(),
 		ProviderID: req.GetProviderId(),
+		Model:      req.GetModel(),
 	})
 	if err != nil {
 		log.WithContext(ctx).Errorf("select policy for request: %v", err)

proxy/internal/middleware/builtin/llm_guardrail/factory.go

@@ -10,11 +10,22 @@ import (
 )
 
 // Config is the JSON-decoded shape accepted by the factory. The
-// runtime path consumes the normalised allowlist; raw config is not
+// runtime path consumes the normalised allowlists; raw config is not
 // retained beyond construction.
 type Config struct {
-	ModelAllowlist []string      `json:"model_allowlist"`
-	PromptCapture  PromptCapture `json:"prompt_capture"`
+	// ProviderAllowlists maps a resolved provider id (the value llm_router
+	// stamps as KeyLLMResolvedProviderID) to that provider's model allowlist. A
+	// provider present here is restricted to the listed models; a provider
+	// absent is unrestricted. The synthesiser only lists a provider when EVERY
+	// policy authorising it enables an allowlist, so a provider reachable by any
+	// un-guardrailed policy is intentionally absent (unrestricted) here and the
+	// precise per-policy/group decision is left to management. Keeping the gate
+	// per-provider — rather than one account-wide union — is what stops a model
+	// allowlisted for one provider from leaking onto another and stops an
+	// un-guardrailed policy's traffic from being blocked by an unrelated
+	// policy's allowlist.
+	ProviderAllowlists map[string][]string `json:"provider_allowlists,omitempty"`
+	PromptCapture      PromptCapture       `json:"prompt_capture"`
 }
 
 // PromptCapture toggles the optional prompt capture + redaction step
@@ -55,20 +66,28 @@ func isEmptyJSON(raw []byte) bool {
 }
 
 // normaliseConfig lowercases and trims allowlist entries so the runtime
-// match is case-insensitive. Empty entries are dropped.
+// match is case-insensitive. Empty entries are dropped. A provider whose
+// entries all drop out keeps an empty (non-nil) list — an allowlist that
+// permits nothing — which is the intended "deny every model" for that
+// provider, distinct from the provider being absent (unrestricted).
 func normaliseConfig(cfg Config) Config {
-	if len(cfg.ModelAllowlist) == 0 {
+	if len(cfg.ProviderAllowlists) == 0 {
+		cfg.ProviderAllowlists = nil
 		return cfg
 	}
-	cleaned := make([]string, 0, len(cfg.ModelAllowlist))
-	for _, entry := range cfg.ModelAllowlist {
-		n := normaliseModel(entry)
-		if n == "" {
-			continue
+	cleaned := make(map[string][]string, len(cfg.ProviderAllowlists))
+	for provider, models := range cfg.ProviderAllowlists {
+		list := make([]string, 0, len(models))
+		for _, entry := range models {
+			n := normaliseModel(entry)
+			if n == "" {
+				continue
+			}
+			list = append(list, n)
 		}
-		cleaned = append(cleaned, n)
+		cleaned[provider] = list
 	}
-	cfg.ModelAllowlist = cleaned
+	cfg.ProviderAllowlists = cleaned
 	return cfg
 }
 

proxy/internal/middleware/builtin/llm_guardrail/middleware.go

@@ -83,8 +83,9 @@ func (m *Middleware) MutationsSupported() bool { return false }
 // prompt capture only affects the metadata emitted alongside an allow.
 func (m *Middleware) Invoke(_ context.Context, in *middleware.Input) (*middleware.Output, error) {
 	model, modelPresent := lookupMetadata(in.Metadata, middleware.KeyLLMModel)
+	providerID, _ := lookupMetadata(in.Metadata, middleware.KeyLLMResolvedProviderID)
 
-	if denial := m.evaluateAllowlist(model, modelPresent); denial != nil {
+	if denial := m.evaluateAllowlist(providerID, model, modelPresent); denial != nil {
 		return denial, nil
 	}
 
@@ -110,20 +111,38 @@ func (m *Middleware) Invoke(_ context.Context, in *middleware.Input) (*middlewar
 // is a no-op.
 func (m *Middleware) Close() error { return nil }
 
-// evaluateAllowlist returns a deny Output when the configured allowlist
-// rejects the model. A nil return means the request should proceed.
-func (m *Middleware) evaluateAllowlist(model string, modelPresent bool) *middleware.Output {
-	if len(m.cfg.ModelAllowlist) == 0 {
+// evaluateAllowlist returns a deny Output when the allowlist for the request's
+// resolved provider rejects the model. A nil return means the request should
+// proceed. The allowlist is scoped to the provider llm_router resolved: a
+// provider absent from the config is unrestricted, so an un-guardrailed policy's
+// traffic is never caught by an unrelated provider's allowlist.
+func (m *Middleware) evaluateAllowlist(providerID, model string, modelPresent bool) *middleware.Output {
+	if len(m.cfg.ProviderAllowlists) == 0 {
 		return nil
 	}
-	// Fail closed: with an allowlist configured, a request whose model the
-	// upstream parser could not extract (absent or empty) must be denied rather
-	// than allowed. This is what enforces the allowlist for URL/path-routed
-	// providers (Bedrock, Vertex, ...) whose model lives outside the JSON body.
+	// Restrictions exist for this account but the resolved provider is unknown,
+	// so we cannot tell whether this request is bound for a restricted provider.
+	// Fail closed rather than wave it through. In practice llm_router always
+	// stamps the resolved provider before the guardrail runs (or denies first),
+	// so this is a defensive guard, not the common path.
+	if providerID == "" {
+		return denyModel("", denyCodeModelUnknown, denyMessageModelUnknown, denyReasonModelUnknown)
+	}
+	allowlist, restricted := m.cfg.ProviderAllowlists[providerID]
+	if !restricted {
+		// This provider has no allowlist (some authorising policy left it
+		// unrestricted); management owns any per-policy/group decision.
+		return nil
+	}
+	// Fail closed: with an allowlist in effect for this provider, a request whose
+	// model the upstream parser could not extract (absent or empty) must be
+	// denied rather than allowed. This is what enforces the allowlist for
+	// URL/path-routed providers (Bedrock, Vertex, ...) whose model lives outside
+	// the JSON body.
 	if !modelPresent || normaliseModel(model) == "" {
 		return denyModel("", denyCodeModelUnknown, denyMessageModelUnknown, denyReasonModelUnknown)
 	}
-	if m.modelInAllowlist(model) {
+	if modelInAllowlist(allowlist, model) {
 		return nil
 	}
 	return denyModel(model, denyCodeModel, denyMessageModel, denyReasonModel)
@@ -151,14 +170,15 @@ func denyModel(model, code, message, reason string) *middleware.Output {
 	}
 }
 
-// modelInAllowlist reports whether the model matches any allowlist
-// entry under the case-insensitive, trim-tolerant comparison rule.
-func (m *Middleware) modelInAllowlist(model string) bool {
+// modelInAllowlist reports whether the model matches any entry in the supplied
+// (already-normalised) allowlist under the case-insensitive, trim-tolerant
+// comparison rule.
+func modelInAllowlist(allowlist []string, model string) bool {
 	normalised := normaliseModel(model)
 	if normalised == "" {
 		return false
 	}
-	for _, allowed := range m.cfg.ModelAllowlist {
+	for _, allowed := range allowlist {
 		if allowed == normalised {
 			return true
 		}

proxy/internal/middleware/builtin/llm_guardrail/middleware_test.go

@@ -26,6 +26,25 @@ func newInput(meta ...middleware.KV) *middleware.Input {
 	return &middleware.Input{Slot: middleware.SlotOnRequest, Metadata: meta}
 }
 
+const (
+	testProvider  = "prov-1"
+	otherProvider = "prov-2"
+)
+
+// providerCfg builds a Config restricting testProvider to the given models.
+func providerCfg(models ...string) Config {
+	return Config{ProviderAllowlists: map[string][]string{testProvider: models}}
+}
+
+// newInputProvider builds an input that carries a resolved provider id (as
+// llm_router would stamp) plus any extra metadata.
+func newInputProvider(provider string, meta ...middleware.KV) *middleware.Input {
+	all := make([]middleware.KV, 0, len(meta)+1)
+	all = append(all, middleware.KV{Key: middleware.KeyLLMResolvedProviderID, Value: provider})
+	all = append(all, meta...)
+	return &middleware.Input{Slot: middleware.SlotOnRequest, Metadata: all}
+}
+
 func TestMiddlewareIdentity(t *testing.T) {
 	mw := New(Config{})
 	assert.Equal(t, ID, mw.ID(), "middleware ID must be llm_guardrail")
@@ -47,12 +66,12 @@ func TestMiddlewareIdentity(t *testing.T) {
 
 func TestAllowlistEmptyAllowsAnyModel(t *testing.T) {
 	mw := New(Config{})
-	out, err := mw.Invoke(context.Background(), newInput(
+	out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
 		middleware.KV{Key: middleware.KeyLLMModel, Value: "gpt-4o"},
 	))
 	require.NoError(t, err)
 	require.NotNil(t, out)
-	assert.Equal(t, middleware.DecisionAllow, out.Decision, "empty allowlist must allow any model")
+	assert.Equal(t, middleware.DecisionAllow, out.Decision, "no provider allowlists must allow any model")
 	v, ok := metaValue(t, out.Metadata, middleware.KeyLLMPolicyDecision)
 	require.True(t, ok, "decision metadata must be emitted")
 	assert.Equal(t, "allow", v, "decision must be allow")
@@ -62,8 +81,8 @@ func TestAllowlistEmptyAllowsAnyModel(t *testing.T) {
 }
 
 func TestAllowlistMatchAllows(t *testing.T) {
-	mw := New(Config{ModelAllowlist: []string{"gpt-4o", "claude-opus-4"}})
-	out, err := mw.Invoke(context.Background(), newInput(
+	mw := New(providerCfg("gpt-4o", "claude-opus-4"))
+	out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
 		middleware.KV{Key: middleware.KeyLLMModel, Value: "gpt-4o"},
 	))
 	require.NoError(t, err)
@@ -71,8 +90,8 @@ func TestAllowlistMatchAllows(t *testing.T) {
 }
 
 func TestAllowlistMissDenies(t *testing.T) {
-	mw := New(Config{ModelAllowlist: []string{"gpt-4o"}})
-	out, err := mw.Invoke(context.Background(), newInput(
+	mw := New(providerCfg("gpt-4o"))
+	out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
 		middleware.KV{Key: middleware.KeyLLMModel, Value: "claude-opus-4"},
 	))
 	require.NoError(t, err)
@@ -91,10 +110,10 @@ func TestAllowlistMissDenies(t *testing.T) {
 }
 
 func TestAllowlistCaseInsensitive(t *testing.T) {
-	mw := New(Config{ModelAllowlist: []string{"  GPT-4o  ", "Claude-OPUS-4"}})
+	mw := New(providerCfg("  GPT-4o  ", "Claude-OPUS-4"))
 	cases := []string{"gpt-4o", "GPT-4O", "  claude-opus-4 "}
 	for _, model := range cases {
-		out, err := mw.Invoke(context.Background(), newInput(
+		out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
 			middleware.KV{Key: middleware.KeyLLMModel, Value: model},
 		))
 		require.NoError(t, err)
@@ -103,14 +122,15 @@ func TestAllowlistCaseInsensitive(t *testing.T) {
 }
 
 func TestAllowlistMissingModelKeyDenies(t *testing.T) {
-	// Fail closed: with an allowlist configured, a request whose model the
-	// parser could not extract (URL/path-routed providers such as Bedrock or
-	// Vertex whose shape wasn't recognised) must be denied, not allowed.
-	mw := New(Config{ModelAllowlist: []string{"gpt-4o"}})
-	out, err := mw.Invoke(context.Background(), newInput())
+	// Fail closed: with an allowlist in effect for the resolved provider, a
+	// request whose model the parser could not extract (URL/path-routed
+	// providers such as Bedrock or Vertex whose shape wasn't recognised) must be
+	// denied, not allowed.
+	mw := New(providerCfg("gpt-4o"))
+	out, err := mw.Invoke(context.Background(), newInputProvider(testProvider))
 	require.NoError(t, err)
 	require.NotNil(t, out)
-	assert.Equal(t, middleware.DecisionDeny, out.Decision, "absent model must be denied when an allowlist is set")
+	assert.Equal(t, middleware.DecisionDeny, out.Decision, "absent model must be denied when the provider is restricted")
 	assert.Equal(t, 403, out.DenyStatus, "deny status must be 403")
 	require.NotNil(t, out.DenyReason, "deny reason must be populated")
 	assert.Equal(t, "llm_policy.model_unknown", out.DenyReason.Code, "deny code must be model_unknown")
@@ -122,26 +142,72 @@ func TestAllowlistMissingModelKeyDenies(t *testing.T) {
 
 func TestAllowlistEmptyModelValueDenies(t *testing.T) {
 	// A present-but-empty model is as undeterminable as an absent one.
-	mw := New(Config{ModelAllowlist: []string{"gpt-4o"}})
-	out, err := mw.Invoke(context.Background(), newInput(
+	mw := New(providerCfg("gpt-4o"))
+	out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
 		middleware.KV{Key: middleware.KeyLLMModel, Value: "   "},
 	))
 	require.NoError(t, err)
 	require.NotNil(t, out)
-	assert.Equal(t, middleware.DecisionDeny, out.Decision, "empty model must be denied when an allowlist is set")
+	assert.Equal(t, middleware.DecisionDeny, out.Decision, "empty model must be denied when the provider is restricted")
 	require.NotNil(t, out.DenyReason, "deny reason must be populated")
 	assert.Equal(t, "llm_policy.model_unknown", out.DenyReason.Code, "deny code must be model_unknown")
 }
 
 func TestAllowlistEmptyListAllowsMissingModel(t *testing.T) {
-	// Without an allowlist there is nothing to enforce, so a missing model is
-	// still allowed — the fail-closed rule only applies when a list is set.
+	// Without any provider allowlists there is nothing to enforce, so a missing
+	// model is still allowed — the fail-closed rule only applies when a
+	// restriction is in effect.
 	mw := New(Config{})
 	out, err := mw.Invoke(context.Background(), newInput())
 	require.NoError(t, err)
 	assert.Equal(t, middleware.DecisionAllow, out.Decision, "no allowlist must allow even without a model")
 }
 
+func TestUnrestrictedProviderAllowsAnyModel(t *testing.T) {
+	// testProvider is restricted, but the request resolved to otherProvider,
+	// which carries no allowlist. Its traffic must not be caught by the other
+	// provider's restriction — this is the cross-provider-leak / false-deny
+	// guard. Management owns any per-policy/group decision for otherProvider.
+	mw := New(providerCfg("gpt-4o"))
+	out, err := mw.Invoke(context.Background(), newInputProvider(otherProvider,
+		middleware.KV{Key: middleware.KeyLLMModel, Value: "claude-opus-4"},
+	))
+	require.NoError(t, err)
+	assert.Equal(t, middleware.DecisionAllow, out.Decision, "an unrestricted provider must not inherit another provider's allowlist")
+}
+
+func TestPerProviderAllowlistsAreIsolated(t *testing.T) {
+	// gpt-4o is allowed only on testProvider; claude-opus-4 only on
+	// otherProvider. A model allowlisted for one provider must not be usable on
+	// the other — the fail-closed layer never unions allowlists across providers.
+	mw := New(Config{ProviderAllowlists: map[string][]string{
+		testProvider:  {"gpt-4o"},
+		otherProvider: {"claude-opus-4"},
+	}})
+	out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
+		middleware.KV{Key: middleware.KeyLLMModel, Value: "claude-opus-4"},
+	))
+	require.NoError(t, err)
+	assert.Equal(t, middleware.DecisionDeny, out.Decision, "claude-opus-4 is allowed only on otherProvider, not testProvider")
+	require.NotNil(t, out.DenyReason)
+	assert.Equal(t, "llm_policy.model_blocked", out.DenyReason.Code, "cross-provider model must be blocked, not model_unknown")
+}
+
+func TestRestrictionsButNoResolvedProviderFailsClosed(t *testing.T) {
+	// Restrictions exist for the account but the resolved provider id is absent,
+	// so the request cannot be scoped to a provider. Fail closed rather than
+	// wave it through.
+	mw := New(providerCfg("gpt-4o"))
+	out, err := mw.Invoke(context.Background(), newInput(
+		middleware.KV{Key: middleware.KeyLLMModel, Value: "gpt-4o"},
+	))
+	require.NoError(t, err)
+	require.NotNil(t, out)
+	assert.Equal(t, middleware.DecisionDeny, out.Decision, "missing resolved provider must fail closed when restrictions exist")
+	require.NotNil(t, out.DenyReason)
+	assert.Equal(t, "llm_policy.model_unknown", out.DenyReason.Code, "deny code must be model_unknown")
+}
+
 func TestPromptCaptureDisabledEmitsNoPrompt(t *testing.T) {
 	mw := New(Config{})
 	out, err := mw.Invoke(context.Background(), newInput(
@@ -217,8 +283,8 @@ func TestFactoryAcceptsZeroConfigs(t *testing.T) {
 
 func TestFactoryDecodesValidConfig(t *testing.T) {
 	cfg := Config{
-		ModelAllowlist: []string{"gpt-4o"},
-		PromptCapture:  PromptCapture{Enabled: true, RedactPii: true},
+		ProviderAllowlists: map[string][]string{testProvider: {"gpt-4o"}},
+		PromptCapture:      PromptCapture{Enabled: true, RedactPii: true},
 	}
 	raw, err := json.Marshal(cfg)
 	require.NoError(t, err, "marshalling test config must succeed")
@@ -234,15 +300,15 @@ func TestFactoryRejectsMalformedJSON(t *testing.T) {
 }
 
 func TestFactoryNormalisesAllowlist(t *testing.T) {
-	raw := []byte(`{"model_allowlist":["  GPT-4o  ","",""," Claude-3 "]}`)
+	raw := []byte(`{"provider_allowlists":{"prov-1":["  GPT-4o  ","",""," Claude-3 "]}}`)
 	mw, err := Factory{}.New(raw)
 	require.NoError(t, err)
-	out, err := mw.Invoke(context.Background(), newInput(
+	out, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
 		middleware.KV{Key: middleware.KeyLLMModel, Value: "gpt-4o"},
 	))
 	require.NoError(t, err)
 	assert.Equal(t, middleware.DecisionAllow, out.Decision, "factory must lowercase + trim allowlist entries")
-	out2, err := mw.Invoke(context.Background(), newInput(
+	out2, err := mw.Invoke(context.Background(), newInputProvider(testProvider,
 		middleware.KV{Key: middleware.KeyLLMModel, Value: "claude-3"},
 	))
 	require.NoError(t, err)

proxy/internal/middleware/builtin/llm_request_parser/guardrail_allowlist_test.go

@@ -25,10 +25,17 @@ func runParserGuardrail(t *testing.T, url string, body []byte, allowlist []strin
 	})
 	require.NoError(t, err, "parser must not error")
 
-	guard := llm_guardrail.New(llm_guardrail.Config{ModelAllowlist: allowlist})
+	const providerID = "prov-under-test"
+	guard := llm_guardrail.New(llm_guardrail.Config{
+		ProviderAllowlists: map[string][]string{providerID: allowlist},
+	})
+	// The real chain has llm_router stamp the resolved provider id before the
+	// guardrail runs; the parser doesn't, so add it here so the guardrail can
+	// scope the allowlist to this provider.
+	meta := append([]middleware.KV{{Key: middleware.KeyLLMResolvedProviderID, Value: providerID}}, parsed.Metadata...)
 	out, err := guard.Invoke(context.Background(), &middleware.Input{
 		Slot:     middleware.SlotOnRequest,
-		Metadata: parsed.Metadata,
+		Metadata: meta,
 	})
 	require.NoError(t, err, "guardrail must not error")
 	require.NotNil(t, out, "guardrail must return an output")