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.
This commit is contained in:
mlsmaycon
2026-07-24 11:42:33 +00:00
parent c48347c51e
commit 66e48bc933
16 changed files with 967 additions and 113 deletions

View File

@@ -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).

View File

@@ -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}` | |

View File

@@ -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` |