feat: fine grained access control policy

This allows a set of rules to be defined for determining access to resources.

Signed-off-by: Gordon Sim <gsim@redhat.com>

llama_stack/distribution/access_control.py

@@ -1,86 +0,0 @@
-# Copyright (c) Meta Platforms, Inc. and affiliates.
-# All rights reserved.
-#
-# This source code is licensed under the terms described in the LICENSE file in
-# the root directory of this source tree.
-
-from typing import Any
-
-from llama_stack.distribution.datatypes import AccessAttributes
-from llama_stack.log import get_logger
-
-logger = get_logger(__name__, category="core")
-
-
-def check_access(
-    obj_identifier: str,
-    obj_attributes: AccessAttributes | None,
-    user_attributes: dict[str, Any] | None = None,
-) -> bool:
-    """Check if the current user has access to the given object, based on access attributes.
-
-    Access control algorithm:
-    1. If the resource has no access_attributes, access is GRANTED to all authenticated users
-    2. If the user has no attributes, access is DENIED to any object with access_attributes defined
-    3. For each attribute category in the resource's access_attributes:
-       a. If the user lacks that category, access is DENIED
-       b. If the user has the category but none of the required values, access is DENIED
-       c. If the user has at least one matching value in each required category, access is GRANTED
-
-    Example:
-        # Resource requires:
-        access_attributes = AccessAttributes(
-            roles=["admin", "data-scientist"],
-            teams=["ml-team"]
-        )
-
-        # User has:
-        user_attributes = {
-            "roles": ["data-scientist", "engineer"],
-            "teams": ["ml-team", "infra-team"],
-            "projects": ["llama-3"]
-        }
-
-        # Result: Access GRANTED
-        # - User has the "data-scientist" role (matches one of the required roles)
-        # - AND user is part of the "ml-team" (matches the required team)
-        # - The extra "projects" attribute is ignored
-
-    Args:
-        obj_identifier: The identifier of the resource object to check access for
-        obj_attributes: The access attributes of the resource object
-        user_attributes: The attributes of the current user
-
-    Returns:
-        bool: True if access is granted, False if denied
-    """
-    # If object has no access attributes, allow access by default
-    if not obj_attributes:
-        return True
-
-    # If no user attributes, deny access to objects with access control
-    if not user_attributes:
-        return False
-
-    dict_attribs = obj_attributes.model_dump(exclude_none=True)
-    if not dict_attribs:
-        return True
-
-    # Check each attribute category (requires ALL categories to match)
-    # TODO: formalize this into a proper ABAC policy
-    for attr_key, required_values in dict_attribs.items():
-        user_values = user_attributes.get(attr_key, [])
-
-        if not user_values:
-            logger.debug(f"Access denied to {obj_identifier}: missing required attribute category '{attr_key}'")
-            return False
-
-        if not any(val in user_values for val in required_values):
-            logger.debug(
-                f"Access denied to {obj_identifier}: "
-                f"no match for attribute '{attr_key}', required one of {required_values}"
-            )
-            return False
-
-    logger.debug(f"Access granted to {obj_identifier}")
-    return True

llama_stack/distribution/access_control/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.

llama_stack/distribution/access_control/access_control.py

@@ -0,0 +1,175 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from typing import Any, Protocol
+
+from llama_stack.distribution.request_headers import User
+
+from .datatypes import (
+    AccessAttributes,
+    AccessRule,
+    Action,
+    AttributeReference,
+    Condition,
+    Scope,
+)
+
+
+def matches_resource(resource_scope: str, actual_resource: str) -> bool:
+    if resource_scope == actual_resource:
+        return True
+    return resource_scope.endswith("::*") and actual_resource.startswith(resource_scope[:-1])
+
+
+def matches_scope(
+    scope: Scope,
+    action: Action,
+    resource: str,
+    user: str | None,
+) -> bool:
+    if scope.resource and not matches_resource(scope.resource, resource):
+        return False
+    if scope.principal and scope.principal != user:
+        return False
+    return action in scope.actions
+
+
+def user_in_literal(
+    literal: str,
+    user_attributes: dict[str, list[str]] | None,
+) -> bool:
+    for qualifier in ["role::", "team::", "project::", "namespace::"]:
+        if literal.startswith(qualifier):
+            if not user_attributes:
+                return False
+            ref = qualifier.replace("::", "s")
+            if ref in user_attributes:
+                value = literal.removeprefix(qualifier)
+                return value in user_attributes[ref]
+            else:
+                return False
+    return False
+
+
+def user_in(
+    ref: AttributeReference | str,
+    resource_attributes: AccessAttributes | None,
+    user_attributes: dict[str, list[str]] | None,
+) -> bool:
+    if not ref.startswith("resource."):
+        return user_in_literal(ref, user_attributes)
+    name = ref.removeprefix("resource.")
+    required = resource_attributes and getattr(resource_attributes, name)
+    if not required:
+        return True
+    if not user_attributes or name not in user_attributes:
+        return False
+    actual = user_attributes[name]
+    for value in required:
+        if value in actual:
+            return True
+    return False
+
+
+def as_list(obj: Any) -> list[Any]:
+    if isinstance(obj, list):
+        return obj
+    return [obj]
+
+
+def matches_conditions(
+    conditions: list[Condition],
+    resource_attributes: AccessAttributes | None,
+    user_attributes: dict[str, list[str]] | None,
+) -> bool:
+    for condition in conditions:
+        # must match all conditions
+        if not matches_condition(condition, resource_attributes, user_attributes):
+            return False
+    return True
+
+
+def matches_condition(
+    condition: Condition | list[Condition],
+    resource_attributes: AccessAttributes | None,
+    user_attributes: dict[str, list[str]] | None,
+) -> bool:
+    if isinstance(condition, list):
+        return matches_conditions(as_list(condition), resource_attributes, user_attributes)
+    if condition.user_in:
+        for ref in as_list(condition.user_in):
+            # if multiple references are specified, all must match
+            if not user_in(ref, resource_attributes, user_attributes):
+                return False
+        return True
+    if condition.user_not_in:
+        for ref in as_list(condition.user_not_in):
+            # if multiple references are specified, none must match
+            if user_in(ref, resource_attributes, user_attributes):
+                return False
+        return True
+    return True
+
+
+def default_policy() -> list[AccessRule]:
+    # for backwards compatibility, if no rules are provided , assume
+    # full access to all subject to attribute matching rules
+    return [
+        AccessRule(
+            permit=Scope(actions=list(Action)),
+            when=Condition(user_in=list(AttributeReference)),
+        )
+    ]
+
+
+class ProtectedResource(Protocol):
+    type: str
+    identifier: str
+    access_attributes: AccessAttributes
+
+
+def is_action_allowed(
+    policy: list[AccessRule],
+    action: Action,
+    resource: ProtectedResource,
+    user: User | None,
+) -> bool:
+    # If user is not set, assume authentication is not enabled
+    if not user:
+        return True
+
+    if not len(policy):
+        policy = default_policy()
+
+    resource_attributes = AccessAttributes()
+    if hasattr(resource, "access_attributes"):
+        resource_attributes = resource.access_attributes
+    qualified_resource_id = resource.type + "::" + resource.identifier
+    for rule in policy:
+        if rule.forbid and matches_scope(rule.forbid, action, qualified_resource_id, user.principal):
+            if rule.when:
+                if matches_condition(rule.when, resource_attributes, user.attributes):
+                    return False
+            elif rule.unless:
+                if not matches_condition(rule.unless, resource_attributes, user.attributes):
+                    return False
+            else:
+                return False
+        elif rule.permit and matches_scope(rule.permit, action, qualified_resource_id, user.principal):
+            if rule.when:
+                if matches_condition(rule.when, resource_attributes, user.attributes):
+                    return True
+            elif rule.unless:
+                if not matches_condition(rule.unless, resource_attributes, user.attributes):
+                    return True
+            else:
+                return True
+    # assume access is denied unless we find a rule that permits access
+    return False
+
+
+class AccessDeniedError(RuntimeError):
+    pass

llama_stack/distribution/access_control/datatypes.py

@@ -0,0 +1,136 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from enum import Enum
+
+from pydantic import BaseModel, Field, model_validator
+from typing_extensions import Self
+
+
+class AccessAttributes(BaseModel):
+    """Structured representation of user attributes for access control.
+
+    This model defines a structured approach to representing user attributes
+    with common standard categories for access control.
+
+    Standard attribute categories include:
+    - roles: Role-based attributes (e.g., admin, data-scientist)
+    - teams: Team-based attributes (e.g., ml-team, infra-team)
+    - projects: Project access attributes (e.g., llama-3, customer-insights)
+    - namespaces: Namespace-based access control for resource isolation
+    """
+
+    # Standard attribute categories - the minimal set we need now
+    roles: list[str] | None = Field(
+        default=None, description="Role-based attributes (e.g., 'admin', 'data-scientist', 'user')"
+    )
+
+    teams: list[str] | None = Field(default=None, description="Team-based attributes (e.g., 'ml-team', 'nlp-team')")
+
+    projects: list[str] | None = Field(
+        default=None, description="Project-based access attributes (e.g., 'llama-3', 'customer-insights')"
+    )
+
+    namespaces: list[str] | None = Field(
+        default=None, description="Namespace-based access control for resource isolation"
+    )
+
+
+class Action(str, Enum):
+    CREATE = "create"
+    READ = "read"
+    UPDATE = "update"
+    DELETE = "delete"
+
+
+class Scope(BaseModel):
+    principal: str | None = None
+    actions: Action | list[Action]
+    resource: str | None = None
+
+
+def _mutually_exclusive(obj, a: str, b: str):
+    if getattr(obj, a) and getattr(obj, b):
+        raise ValueError(f"{a} and {b} are mutually exclusive")
+
+
+def _require_one_of(obj, a: str, b: str):
+    if not getattr(obj, a) and not getattr(obj, b):
+        raise ValueError(f"on of {a} or {b} is required")
+
+
+class AttributeReference(str, Enum):
+    RESOURCE_ROLES = "resource.roles"
+    RESOURCE_TEAMS = "resource.teams"
+    RESOURCE_PROJECTS = "resource.projects"
+    RESOURCE_NAMESPACES = "resource.namespaces"
+
+
+class Condition(BaseModel):
+    user_in: AttributeReference | list[AttributeReference] | str | None = None
+    user_not_in: AttributeReference | list[AttributeReference] | str | None = None
+
+
+class AccessRule(BaseModel):
+    """Access rule based loosely on cedar policy language
+
+    A rule defines a list of action either to permit or to forbid. It may specify a
+    principal or a resource that must match for the rule to take effect. The resource
+    to match should be specified in the form of a type qualified identifier, e.g.
+    model::my-model or vector_db::some-db, or a wildcard for all resources of a type,
+    e.g. model::*. If the principal or resource are not specified, they will match all
+    requests.
+
+    A rule may also specify a condition, either a 'when' or an 'unless', with additional
+    constraints as to where the rule applies. The constraints at present are whether the
+    user requesting access is in or not in some set. This set can either be a particular
+    set of attributes on the resource e.g. resource.roles or a literal value of some
+    notion of group, e.g. role::admin or namespace::foo.
+
+    Rules are tested in order to find a match. If a match is found, the request is
+    permitted or forbidden depending on the type of rule. If no match is found, the
+    request is denied. If no rules are specified, a rule that allows any action as
+    long as the resource attributes match the user attributes is added
+    (i.e. the previous behaviour is the default).
+
+    Some examples in yaml:
+
+    - permit:
+      principal: user-1
+      actions: [create, read, delete]
+      resource: model::*
+      description: user-1 has full access to all models
+    - permit:
+      principal: user-2
+      actions: [read]
+      resource: model::model-1
+      description: user-2 has read access to model-1 only
+    - permit:
+      actions: [read]
+      when:
+        user_in: resource.namespaces
+      description: any user has read access to any resource with matching attributes
+    - forbid:
+      actions: [create, read, delete]
+      resource: vector_db::*
+      unless:
+        user_in: role::admin
+      description: only user with admin role can use vector_db resources
+
+    """
+
+    permit: Scope | None = None
+    forbid: Scope | None = None
+    when: Condition | list[Condition] | None = None
+    unless: Condition | list[Condition] | None = None
+    description: str | None = None
+
+    @model_validator(mode="after")
+    def validate_rule_format(self) -> Self:
+        _require_one_of(self, "permit", "forbid")
+        _mutually_exclusive(self, "permit", "forbid")
+        _mutually_exclusive(self, "when", "unless")
+        return self

llama_stack/distribution/datatypes.py

@@ -24,6 +24,7 @@ from llama_stack.apis.shields import Shield, ShieldInput
 from llama_stack.apis.tools import Tool, ToolGroup, ToolGroupInput, ToolRuntime
 from llama_stack.apis.vector_dbs import VectorDB, VectorDBInput
 from llama_stack.apis.vector_io import VectorIO
+from llama_stack.distribution.access_control.datatypes import AccessAttributes, AccessRule
 from llama_stack.providers.datatypes import Api, ProviderSpec
 from llama_stack.providers.utils.kvstore.config import KVStoreConfig, SqliteKVStoreConfig
 from llama_stack.providers.utils.sqlstore.sqlstore import SqlStoreConfig
@@ -35,35 +36,6 @@ LLAMA_STACK_RUN_CONFIG_VERSION = "2"
 RoutingKey = str | list[str]
 
 
-class AccessAttributes(BaseModel):
-    """Structured representation of user attributes for access control.
-
-    This model defines a structured approach to representing user attributes
-    with common standard categories for access control.
-
-    Standard attribute categories include:
-    - roles: Role-based attributes (e.g., admin, data-scientist)
-    - teams: Team-based attributes (e.g., ml-team, infra-team)
-    - projects: Project access attributes (e.g., llama-3, customer-insights)
-    - namespaces: Namespace-based access control for resource isolation
-    """
-
-    # Standard attribute categories - the minimal set we need now
-    roles: list[str] | None = Field(
-        default=None, description="Role-based attributes (e.g., 'admin', 'data-scientist', 'user')"
-    )
-
-    teams: list[str] | None = Field(default=None, description="Team-based attributes (e.g., 'ml-team', 'nlp-team')")
-
-    projects: list[str] | None = Field(
-        default=None, description="Project-based access attributes (e.g., 'llama-3', 'customer-insights')"
-    )
-
-    namespaces: list[str] | None = Field(
-        default=None, description="Namespace-based access control for resource isolation"
-    )
-
-
 class ResourceWithACL(Resource):
     """Extension of Resource that adds attribute-based access control capabilities.
 
@@ -234,6 +206,7 @@ class AuthenticationConfig(BaseModel):
         ...,
         description="Provider-specific configuration",
     )
+    access_policy: list[AccessRule] = Field(default=[], description="Rules for determining access to resources")
 
 
 class AuthenticationRequiredError(Exception):

llama_stack/distribution/request_headers.py

@@ -18,15 +18,23 @@ log = logging.getLogger(__name__)
 PROVIDER_DATA_VAR = contextvars.ContextVar("provider_data", default=None)
 
 
+class User:
+    principal: str
+    # further attributes that may be used for access control decisions
+    attributes: dict[str, list[str]]
+
+    def __init__(self, principal: str, attributes: dict[str, list[str]]):
+        self.principal = principal
+        self.attributes = attributes
+
+
 class RequestProviderDataContext(AbstractContextManager):
     """Context manager for request provider data"""
 
-    def __init__(
-        self, provider_data: dict[str, Any] | None = None, auth_attributes: dict[str, list[str]] | None = None
-    ):
+    def __init__(self, provider_data: dict[str, Any] | None = None, user: User | None = None):
         self.provider_data = provider_data or {}
-        if auth_attributes:
-            self.provider_data["__auth_attributes"] = auth_attributes
+        if user:
+            self.provider_data["__authenticated_user"] = user
 
         self.token = None
 
@@ -95,9 +103,9 @@ def request_provider_data_context(
     return RequestProviderDataContext(provider_data, auth_attributes)
 
 
-def get_auth_attributes() -> dict[str, list[str]] | None:
+def get_authenticated_user() -> User | None:
     """Helper to retrieve auth attributes from the provider data context"""
     provider_data = PROVIDER_DATA_VAR.get()
     if not provider_data:
         return None
-    return provider_data.get("__auth_attributes")
+    return provider_data.get("__authenticated_user")

llama_stack/distribution/resolver.py

@@ -28,6 +28,7 @@ from llama_stack.apis.vector_dbs import VectorDBs
 from llama_stack.apis.vector_io import VectorIO
 from llama_stack.distribution.client import get_client_impl
 from llama_stack.distribution.datatypes import (
+    AccessRule,
     AutoRoutedProviderSpec,
     Provider,
     RoutingTableProviderSpec,
@@ -118,6 +119,7 @@ async def resolve_impls(
     run_config: StackRunConfig,
     provider_registry: ProviderRegistry,
     dist_registry: DistributionRegistry,
+    policy: list[AccessRule],
 ) -> dict[Api, Any]:
     """
     Resolves provider implementations by:
@@ -140,7 +142,7 @@ async def resolve_impls(
 
     sorted_providers = sort_providers_by_deps(providers_with_specs, run_config)
 
-    return await instantiate_providers(sorted_providers, router_apis, dist_registry, run_config)
+    return await instantiate_providers(sorted_providers, router_apis, dist_registry, run_config, policy)
 
 
 def specs_for_autorouted_apis(apis_to_serve: list[str] | set[str]) -> dict[str, dict[str, ProviderWithSpec]]:
@@ -247,6 +249,7 @@ async def instantiate_providers(
     router_apis: set[Api],
     dist_registry: DistributionRegistry,
     run_config: StackRunConfig,
+    policy: list[AccessRule],
 ) -> dict:
     """Instantiates providers asynchronously while managing dependencies."""
     impls: dict[Api, Any] = {}
@@ -261,7 +264,7 @@ async def instantiate_providers(
         if isinstance(provider.spec, RoutingTableProviderSpec):
             inner_impls = inner_impls_by_provider_id[f"inner-{provider.spec.router_api.value}"]
 
-        impl = await instantiate_provider(provider, deps, inner_impls, dist_registry, run_config)
+        impl = await instantiate_provider(provider, deps, inner_impls, dist_registry, run_config, policy)
 
         if api_str.startswith("inner-"):
             inner_impls_by_provider_id[api_str][provider.provider_id] = impl
@@ -312,6 +315,7 @@ async def instantiate_provider(
     inner_impls: dict[str, Any],
     dist_registry: DistributionRegistry,
     run_config: StackRunConfig,
+    policy: list[AccessRule],
 ):
     provider_spec = provider.spec
     if not hasattr(provider_spec, "module"):
@@ -336,13 +340,15 @@ async def instantiate_provider(
         method = "get_routing_table_impl"
 
         config = None
-        args = [provider_spec.api, inner_impls, deps, dist_registry]
+        args = [provider_spec.api, inner_impls, deps, dist_registry, policy]
     else:
         method = "get_provider_impl"
 
         config_type = instantiate_class_type(provider_spec.config_class)
         config = config_type(**provider.config)
         args = [config, deps]
+        if "policy" in inspect.signature(getattr(module, method)).parameters:
+            args.append(policy)
 
     fn = getattr(module, method)
     impl = await fn(*args)

llama_stack/distribution/routers/__init__.py

@@ -6,7 +6,7 @@
 
 from typing import Any
 
-from llama_stack.distribution.datatypes import RoutedProtocol
+from llama_stack.distribution.datatypes import AccessRule, RoutedProtocol
 from llama_stack.distribution.stack import StackRunConfig
 from llama_stack.distribution.store import DistributionRegistry
 from llama_stack.providers.datatypes import Api, RoutingTable
@@ -18,6 +18,7 @@ async def get_routing_table_impl(
     impls_by_provider_id: dict[str, RoutedProtocol],
     _deps,
     dist_registry: DistributionRegistry,
+    policy: list[AccessRule],
 ) -> Any:
     from ..routing_tables.benchmarks import BenchmarksRoutingTable
     from ..routing_tables.datasets import DatasetsRoutingTable
@@ -40,7 +41,7 @@ async def get_routing_table_impl(
     if api.value not in api_to_tables:
         raise ValueError(f"API {api.value} not found in router map")
 
-    impl = api_to_tables[api.value](impls_by_provider_id, dist_registry)
+    impl = api_to_tables[api.value](impls_by_provider_id, dist_registry, policy)
     await impl.initialize()
     return impl
 

llama_stack/distribution/routing_tables/common.py

@@ -8,14 +8,15 @@ from typing import Any
 
 from llama_stack.apis.resource import ResourceType
 from llama_stack.apis.scoring_functions import ScoringFn
-from llama_stack.distribution.access_control import check_access
+from llama_stack.distribution.access_control.access_control import AccessDeniedError, is_action_allowed
 from llama_stack.distribution.datatypes import (
     AccessAttributes,
+    AccessRule,
     RoutableObject,
     RoutableObjectWithProvider,
     RoutedProtocol,
 )
-from llama_stack.distribution.request_headers import get_auth_attributes
+from llama_stack.distribution.request_headers import get_authenticated_user
 from llama_stack.distribution.store import DistributionRegistry
 from llama_stack.log import get_logger
 from llama_stack.providers.datatypes import Api, RoutingTable
@@ -73,9 +74,11 @@ class CommonRoutingTableImpl(RoutingTable):
         self,
         impls_by_provider_id: dict[str, RoutedProtocol],
         dist_registry: DistributionRegistry,
+        policy: list[AccessRule],
     ) -> None:
         self.impls_by_provider_id = impls_by_provider_id
         self.dist_registry = dist_registry
+        self.policy = policy
 
     async def initialize(self) -> None:
         async def add_objects(objs: list[RoutableObjectWithProvider], provider_id: str, cls) -> None:
@@ -166,13 +169,15 @@ class CommonRoutingTableImpl(RoutingTable):
             return None
 
         # Check if user has permission to access this object
-        if not check_access(obj.identifier, getattr(obj, "access_attributes", None), get_auth_attributes()):
-            logger.debug(f"Access denied to {type} '{identifier}' based on attribute mismatch")
+        if not is_action_allowed(self.policy, "read", obj, get_authenticated_user()):
+            logger.debug(f"Access denied to {type} '{identifier}'")
             return None
 
         return obj
 
     async def unregister_object(self, obj: RoutableObjectWithProvider) -> None:
+        if not is_action_allowed(self.policy, "delete", obj, get_authenticated_user()):
+            raise AccessDeniedError()
         await self.dist_registry.delete(obj.type, obj.identifier)
         await unregister_object_from_provider(obj, self.impls_by_provider_id[obj.provider_id])
 
@@ -187,11 +192,12 @@ class CommonRoutingTableImpl(RoutingTable):
         p = self.impls_by_provider_id[obj.provider_id]
 
         # If object supports access control but no attributes set, use creator's attributes
-        if not obj.access_attributes:
-            creator_attributes = get_auth_attributes()
-            if creator_attributes:
-                obj.access_attributes = AccessAttributes(**creator_attributes)
-                logger.info(f"Setting access attributes for {obj.type} '{obj.identifier}' based on creator's identity")
+        creator = get_authenticated_user()
+        if not is_action_allowed(self.policy, "create", obj, creator):
+            raise AccessDeniedError()
+        if creator and creator.attributes:
+            obj.access_attributes = AccessAttributes(**creator.attributes)
+            logger.info(f"Setting access attributes for {obj.type} '{obj.identifier}' based on creator's identity")
 
         registered_obj = await register_object_with_provider(obj, p)
         # TODO: This needs to be fixed for all APIs once they return the registered object
@@ -210,9 +216,7 @@ class CommonRoutingTableImpl(RoutingTable):
         # Apply attribute-based access control filtering
         if filtered_objs:
             filtered_objs = [
-                obj
-                for obj in filtered_objs
-                if check_access(obj.identifier, getattr(obj, "access_attributes", None), get_auth_attributes())
+                obj for obj in filtered_objs if is_action_allowed(self.policy, "read", obj, get_authenticated_user())
             ]
 
         return filtered_objs

llama_stack/distribution/server/server.py

@@ -30,10 +30,7 @@ from pydantic import BaseModel, ValidationError
 
 from llama_stack.distribution.datatypes import AuthenticationRequiredError, LoggingConfig, StackRunConfig
 from llama_stack.distribution.distribution import builtin_automatically_routed_apis
-from llama_stack.distribution.request_headers import (
-    PROVIDER_DATA_VAR,
-    request_provider_data_context,
-)
+from llama_stack.distribution.request_headers import PROVIDER_DATA_VAR, User, request_provider_data_context
 from llama_stack.distribution.resolver import InvalidProviderError
 from llama_stack.distribution.server.endpoints import (
     find_matching_endpoint,
@@ -213,11 +210,13 @@ def create_dynamic_typed_route(func: Any, method: str, route: str):
     async def endpoint(request: Request, **kwargs):
         # Get auth attributes from the request scope
         user_attributes = request.scope.get("user_attributes", {})
+        principal = request.scope.get("principal", "")
+        user = User(principal, user_attributes)
 
         await log_request_pre_validation(request)
 
         # Use context manager with both provider data and auth attributes
-        with request_provider_data_context(request.headers, user_attributes):
+        with request_provider_data_context(request.headers, user):
             is_streaming = is_streaming_request(func.__name__, request, **kwargs)
 
             try:

llama_stack/distribution/stack.py

@@ -223,7 +223,10 @@ async def construct_stack(
     run_config: StackRunConfig, provider_registry: ProviderRegistry | None = None
 ) -> dict[Api, Any]:
     dist_registry, _ = await create_dist_registry(run_config.metadata_store, run_config.image_name)
-    impls = await resolve_impls(run_config, provider_registry or get_provider_registry(run_config), dist_registry)
+    policy = run_config.server.auth.access_policy if run_config.server.auth else []
+    impls = await resolve_impls(
+        run_config, provider_registry or get_provider_registry(run_config), dist_registry, policy
+    )
 
     # Add internal implementations after all other providers are resolved
     add_internal_implementations(impls, run_config)