feat(server): add attribute based access control for resources (#1703)

This PR introduces a way to implement Attribute Based Access Control
(ABAC) for the Llama Stack server.

The rough design is:
- https://github.com/meta-llama/llama-stack/pull/1626 added a way for
the Llama Stack server to query an authenticator
- We build upon that and expect "access attributes" as part of the
response. These attributes indicate the scopes available for the
request.
- We use these attributes to perform access control for registered
resources as well as for constructing the default access control
policies for newly created resources.
- By default, if you support authentication but don't return access
attributes, we will add a unique namespace pointing to the API_KEY. That
way, all resources by default will be scoped to API_KEYs.

An important aspect of this design is that Llama Stack stays out of the
business of credential management or the CRUD for attributes. How you
manage your namespaces or projects is entirely up to you. The design
only implements access control checks for the metadata / book-keeping
information that the Stack tracks.

### Limitations

- Currently, read vs. write vs. admin permissions aren't made explicit,
but this can be easily extended by adding appropriate attributes to the
`AccessAttributes` data structure.
- This design does not apply to agent instances since they are not
considered resources the Stack knows about. Agent instances are
completely within the scope of the Agents API provider.

### Test Plan

Added unit tests, existing integration tests
This commit is contained in:
Ashwin Bharambe 2025-03-19 21:28:52 -07:00 committed by GitHub
parent c4e1b8d094
commit 01a25d9744
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
10 changed files with 890 additions and 45 deletions

View file

@ -5,16 +5,118 @@
# the root directory of this source tree.
import json
from typing import Dict, List, Optional
from urllib.parse import parse_qs
import httpx
from pydantic import BaseModel, Field
from llama_stack.distribution.datatypes import AccessAttributes
from llama_stack.log import get_logger
logger = get_logger(name=__name__, category="auth")
class AuthRequestContext(BaseModel):
path: str = Field(description="The path of the request being authenticated")
headers: Dict[str, str] = Field(description="HTTP headers from the original request (excluding Authorization)")
params: Dict[str, List[str]] = Field(
description="Query parameters from the original request, parsed as dictionary of lists"
)
class AuthRequest(BaseModel):
api_key: str = Field(description="The API key extracted from the Authorization header")
request: AuthRequestContext = Field(description="Context information about the request being authenticated")
class AuthResponse(BaseModel):
"""The format of the authentication response from the auth endpoint."""
access_attributes: Optional[AccessAttributes] = Field(
default=None,
description="""
Structured user attributes for attribute-based access control.
These attributes determine which resources the user can access.
The model provides standard categories like "roles", "teams", "projects", and "namespaces".
Each attribute category contains a list of values that the user has for that category.
During access control checks, these values are compared against resource requirements.
Example with standard categories:
```json
{
"roles": ["admin", "data-scientist"],
"teams": ["ml-team"],
"projects": ["llama-3"],
"namespaces": ["research"]
}
```
""",
)
message: Optional[str] = Field(
default=None, description="Optional message providing additional context about the authentication result."
)
class AuthenticationMiddleware:
"""Middleware that authenticates requests using an external auth endpoint.
This middleware:
1. Extracts the Bearer token from the Authorization header
2. Sends it to the configured auth endpoint along with request details
3. Validates the response and extracts user attributes
4. Makes these attributes available to the route handlers for access control
Authentication Request Format:
```json
{
"api_key": "the-api-key-extracted-from-auth-header",
"request": {
"path": "/models/list",
"headers": {
"content-type": "application/json",
"user-agent": "..."
// All headers except Authorization
},
"params": {
"limit": ["100"],
"offset": ["0"]
// Query parameters as key -> list of values
}
}
}
```
Expected Auth Endpoint Response Format:
```json
{
"access_attributes": { // Structured attribute format
"roles": ["admin", "user"],
"teams": ["ml-team", "nlp-team"],
"projects": ["llama-3", "project-x"],
"namespaces": ["research"]
},
"message": "Optional message about auth result"
}
```
Attribute-Based Access Control:
The attributes returned by the auth endpoint are used to determine which
resources the user can access. Resources can specify required attributes
using the access_attributes field. For a user to access a resource:
1. All attribute categories specified in the resource must be present in the user's attributes
2. For each category, the user must have at least one matching value
If the auth endpoint doesn't return any attributes, the user will only be able to
access resources that don't have access_attributes defined.
"""
def __init__(self, app, auth_endpoint):
self.app = app
self.auth_endpoint = auth_endpoint
@ -32,25 +134,57 @@ class AuthenticationMiddleware:
path = scope.get("path", "")
request_headers = {k.decode(): v.decode() for k, v in headers.items()}
# Remove sensitive headers
if "authorization" in request_headers:
del request_headers["authorization"]
query_string = scope.get("query_string", b"").decode()
params = parse_qs(query_string)
auth_data = {
"api_key": api_key,
"request": {
"path": path,
"headers": request_headers,
"params": params,
},
}
# Build the auth request model
auth_request = AuthRequest(
api_key=api_key,
request=AuthRequestContext(
path=path,
headers=request_headers,
params=params,
),
)
# Validate with authentication endpoint
try:
async with httpx.AsyncClient() as client:
response = await client.post(self.auth_endpoint, json=auth_data)
response = await client.post(
self.auth_endpoint,
json=auth_request.model_dump(),
timeout=10.0, # Add a reasonable timeout
)
if response.status_code != 200:
logger.warning(f"Authentication failed: {response.status_code}")
return await self._send_auth_error(send, "Authentication failed")
# Parse and validate the auth response
try:
response_data = response.json()
auth_response = AuthResponse(**response_data)
# Store attributes in request scope for access control
if auth_response.access_attributes:
user_attributes = auth_response.access_attributes.model_dump(exclude_none=True)
else:
logger.warning("No access attributes, setting namespace to api_key by default")
user_attributes = {
"namespaces": [api_key],
}
scope["user_attributes"] = user_attributes
logger.debug(f"Authentication successful: {len(user_attributes)} attributes")
except Exception:
logger.exception("Error parsing authentication response")
return await self._send_auth_error(send, "Invalid authentication response format")
except httpx.TimeoutException:
logger.exception("Authentication request timed out")
return await self._send_auth_error(send, "Authentication service timeout")
except Exception:
logger.exception("Error during authentication")
return await self._send_auth_error(send, "Authentication service error")

View file

@ -179,8 +179,11 @@ async def sse_generator(event_gen):
def create_dynamic_typed_route(func: Any, method: str, route: str):
async def endpoint(request: Request, **kwargs):
# Use context manager for request provider data
with request_provider_data_context(request.headers):
# Get auth attributes from the request scope
user_attributes = request.scope.get("user_attributes", {})
# Use context manager with both provider data and auth attributes
with request_provider_data_context(request.headers, user_attributes):
is_streaming = is_streaming_request(func.__name__, request, **kwargs)
try: