mirror of
https://github.com/BerriAI/litellm.git
synced 2026-11-15 09:12:18 +00:00
Merge b40e6ad4bf into 6254860560
This commit is contained in:
commit
a0d7ace4a6
2 changed files with 227 additions and 13 deletions
|
|
@ -2337,6 +2337,56 @@ async def test_custom_code_guardrail(
|
|||
)
|
||||
|
||||
|
||||
_GUARDRAIL_METADATA_ADAPTER: Final = TypeAdapter(Mapping[str, object])
|
||||
|
||||
|
||||
def _metadata_fields(value: object) -> Mapping[str, object]:
|
||||
try:
|
||||
return _GUARDRAIL_METADATA_ADAPTER.validate_python(value)
|
||||
except ValidationError:
|
||||
return {}
|
||||
|
||||
|
||||
_CALLER_OWNED_KEY_FIELDS: Final[frozenset[str]] = frozenset({"user_api_key_end_user_id", "user_api_key_request_route"})
|
||||
|
||||
_RESERVED_IDENTITY_FIELDS: Final[frozenset[str]] = frozenset(
|
||||
{"requester_ip_address", "user_agent", "agent_id", "actor_agent_id"}
|
||||
)
|
||||
|
||||
|
||||
def _is_key_identity_field(key: str) -> bool:
|
||||
"""Whether a metadata field attributes the authenticated caller rather than the request.
|
||||
|
||||
Pre-call processing owns the whole ``user_api_key_`` namespace, except the two fields that
|
||||
describe the caller's own request: the end user it is made for, and the route it came from.
|
||||
A guardrail needs both to decide, and on ``/apply_guardrail`` the proxy's own values say
|
||||
nothing about the call the caller is asking about.
|
||||
|
||||
Attribution also reaches outside that namespace, so the fields pre-call processing fills in
|
||||
from the key and the connection are reserved by name as well.
|
||||
"""
|
||||
if key in _CALLER_OWNED_KEY_FIELDS:
|
||||
return False
|
||||
return key in _RESERVED_IDENTITY_FIELDS or key == "user_api_key" or key.startswith("user_api_key_")
|
||||
|
||||
|
||||
def _guardrail_request_metadata(
|
||||
caller: object, proxy: object
|
||||
) -> dict[str, object]: # mutable-ok: guardrails annotate the metadata they are handed
|
||||
"""Caller metadata with the identity fields taken from the authenticated key.
|
||||
|
||||
The result is handed to the guardrail as ``request_data["metadata"]``, which providers
|
||||
write their own detection details into, so it cannot be a read-only mapping.
|
||||
"""
|
||||
caller_fields: Final = tuple(
|
||||
(key, value) for key, value in _metadata_fields(caller).items() if not _is_key_identity_field(key)
|
||||
)
|
||||
key_identity: Final = tuple(
|
||||
(key, value) for key, value in _metadata_fields(proxy).items() if _is_key_identity_field(key)
|
||||
)
|
||||
return dict((*caller_fields, *key_identity))
|
||||
|
||||
|
||||
def _execution_timeout_response(timeout: float) -> TestCustomCodeGuardrailResponse:
|
||||
return TestCustomCodeGuardrailResponse(
|
||||
success=False,
|
||||
|
|
@ -2503,6 +2553,14 @@ async def apply_guardrail(
|
|||
|
||||
response_handler: Final = GUARDRAIL_RESPONSE_HANDLERS[request.response_format]
|
||||
request_data: Final = dict(response_handler.prepare_request_data(request))
|
||||
caller_metadata: Final = request_data.get("metadata")
|
||||
metadata: Final = _guardrail_request_metadata(
|
||||
caller_metadata, _GUARDRAIL_METADATA_ADAPTER.validate_python(data).get("metadata")
|
||||
)
|
||||
if caller_metadata is not None or metadata:
|
||||
request_data["metadata"] = metadata
|
||||
else:
|
||||
request_data.pop("metadata", None)
|
||||
input_type: Final = response_handler.resolve_input_type(request, active_guardrail)
|
||||
guardrailed_inputs: Final = await active_guardrail.apply_guardrail(
|
||||
inputs={"texts": [request.text]},
|
||||
|
|
|
|||
|
|
@ -4,7 +4,7 @@ from collections.abc import Callable
|
|||
from datetime import datetime
|
||||
from types import MappingProxyType
|
||||
from typing import Dict, Final, List, Literal, Optional
|
||||
from unittest.mock import AsyncMock
|
||||
from unittest.mock import AsyncMock, Mock
|
||||
|
||||
import httpx
|
||||
import pytest
|
||||
|
|
@ -25,6 +25,7 @@ from litellm.proxy.guardrails.guardrail_endpoints import (
|
|||
TestCustomCodeGuardrailRequest,
|
||||
UpdateGuardrailRequest,
|
||||
_get_guardrails_list_response,
|
||||
_guardrail_request_metadata,
|
||||
apply_guardrail,
|
||||
approve_guardrail_submission,
|
||||
create_guardrail,
|
||||
|
|
@ -2169,27 +2170,29 @@ async def test_apply_guardrail_invokes_logging_pipeline(mocker):
|
|||
}
|
||||
|
||||
|
||||
def _patch_apply_guardrail_env(mocker, guardrail_result):
|
||||
mock_guardrail = mocker.Mock()
|
||||
def _patch_apply_guardrail_env(
|
||||
mocker: MockerFixture, guardrail_result: object, processed_data: dict[str, object] | None = None
|
||||
) -> Mock:
|
||||
mock_guardrail: Final = mocker.Mock()
|
||||
mock_guardrail.apply_guardrail = AsyncMock(return_value=guardrail_result)
|
||||
|
||||
mock_registry = mocker.Mock()
|
||||
mock_registry: Final = mocker.Mock()
|
||||
mock_registry.get_initialized_guardrail_callback.return_value = mock_guardrail
|
||||
mocker.patch("litellm.proxy.guardrails.guardrail_endpoints.GUARDRAIL_REGISTRY", mock_registry)
|
||||
|
||||
mock_logging_obj = mocker.Mock()
|
||||
mock_logging_obj: Final = mocker.Mock()
|
||||
mock_logging_obj.async_success_handler = AsyncMock()
|
||||
mock_logging_obj.model_call_details = {}
|
||||
mock_processor = mocker.Mock()
|
||||
mock_processor: Final = mocker.Mock()
|
||||
mock_processor.common_processing_pre_call_logic = AsyncMock(
|
||||
return_value=({"guardrail_name": "test-guardrail"}, mock_logging_obj)
|
||||
return_value=(processed_data or {"guardrail_name": "test-guardrail"}, mock_logging_obj)
|
||||
)
|
||||
mocker.patch(
|
||||
"litellm.proxy.common_request_processing.ProxyBaseLLMRequestProcessing",
|
||||
return_value=mock_processor,
|
||||
)
|
||||
|
||||
mock_proxy_logging = mocker.Mock()
|
||||
mock_proxy_logging: Final = mocker.Mock()
|
||||
mock_proxy_logging.post_call_success_hook = AsyncMock()
|
||||
mocker.patch("litellm.proxy.proxy_server.proxy_logging_obj", mock_proxy_logging)
|
||||
mocker.patch("litellm.proxy.proxy_server.general_settings", {})
|
||||
|
|
@ -2483,21 +2486,174 @@ async def test_apply_guardrail_forwards_metadata_and_messages_together(mocker):
|
|||
)
|
||||
|
||||
|
||||
class TestGuardrailRequestMetadata:
|
||||
"""What a guardrail is told about the caller on /apply_guardrail."""
|
||||
|
||||
def test_the_authenticated_key_identity_replaces_the_caller_version(self) -> None:
|
||||
metadata: Final = _guardrail_request_metadata(
|
||||
{"forbidden_topics": ["tax"], "user_api_key_alias": "someone-else"},
|
||||
{"user_api_key_alias": "billing-app", "user_api_key_user_id": "u-1"},
|
||||
)
|
||||
assert metadata == {
|
||||
"forbidden_topics": ["tax"],
|
||||
"user_api_key_alias": "billing-app",
|
||||
"user_api_key_user_id": "u-1",
|
||||
}
|
||||
|
||||
def test_the_bare_key_hash_cannot_come_from_the_caller(self) -> None:
|
||||
metadata: Final = _guardrail_request_metadata({"user_api_key": "spoof"}, {"user_api_key": "hashed"})
|
||||
assert metadata == {"user_api_key": "hashed"}
|
||||
|
||||
def test_the_caller_owns_the_end_user_and_the_route(self) -> None:
|
||||
"""Both describe the call the caller is asking about, not the key it authenticated with."""
|
||||
metadata: Final = _guardrail_request_metadata(
|
||||
{"user_api_key_end_user_id": "customer-42", "user_api_key_request_route": "/v1/chat/completions"},
|
||||
{"user_api_key_request_route": "/apply_guardrail"},
|
||||
)
|
||||
assert metadata == {
|
||||
"user_api_key_end_user_id": "customer-42",
|
||||
"user_api_key_request_route": "/v1/chat/completions",
|
||||
}
|
||||
|
||||
def test_identity_fields_outside_the_key_namespace_are_reserved_too(self) -> None:
|
||||
"""The proxy sets who is calling and from where. A caller cannot claim another's."""
|
||||
metadata: Final = _guardrail_request_metadata(
|
||||
{
|
||||
"requester_ip_address": "203.0.113.9",
|
||||
"user_agent": "forged/1.0",
|
||||
"agent_id": "someone-elses-agent",
|
||||
"topic": "tax",
|
||||
},
|
||||
{
|
||||
"requester_ip_address": "198.51.100.4",
|
||||
"user_agent": "litellm-cli/2.3",
|
||||
"agent_id": "billing-agent",
|
||||
},
|
||||
)
|
||||
assert metadata == {
|
||||
"topic": "tax",
|
||||
"requester_ip_address": "198.51.100.4",
|
||||
"user_agent": "litellm-cli/2.3",
|
||||
"agent_id": "billing-agent",
|
||||
}
|
||||
|
||||
def test_a_guardrail_can_annotate_what_it_is_handed(self) -> None:
|
||||
"""OpenAI moderation and Presidio write their detection details into this mapping."""
|
||||
metadata: Final = _guardrail_request_metadata({}, {"user_api_key_alias": "billing-app"})
|
||||
metadata["moderation_model"] = "omni-moderation-latest"
|
||||
assert metadata == {"user_api_key_alias": "billing-app", "moderation_model": "omni-moderation-latest"}
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_apply_guardrail_omits_metadata_when_not_sent(mocker):
|
||||
"""Without metadata, request_data stays empty (backward-compatible)."""
|
||||
mock_guardrail = _patch_apply_guardrail_env(mocker, {"texts": ["ok"]})
|
||||
async def test_apply_guardrail_replaces_caller_identity_with_the_authenticated_key(mocker):
|
||||
"""Identity fields come from the proxy's own sanitized metadata, never from
|
||||
the caller, so a request cannot impersonate another key or probe its policy."""
|
||||
mock_guardrail = _patch_apply_guardrail_env(
|
||||
mocker,
|
||||
{"texts": ["ok"]},
|
||||
processed_data={
|
||||
"guardrail_name": "test-guardrail",
|
||||
"metadata": {
|
||||
"route": "/apply_guardrail",
|
||||
"user_api_key_alias": "billing-app",
|
||||
"user_api_key_user_id": "u-1",
|
||||
},
|
||||
},
|
||||
)
|
||||
|
||||
request = ApplyGuardrailRequest(
|
||||
guardrail_name="test-guardrail",
|
||||
text="hello",
|
||||
metadata={
|
||||
"forbidden_topics": ["tax"],
|
||||
"user_api_key_alias": "someone-else",
|
||||
"user_api_key_user_email": "victim@example.com",
|
||||
},
|
||||
)
|
||||
await apply_guardrail(
|
||||
fastapi_request=mocker.Mock(),
|
||||
request=request,
|
||||
user_api_key_dict=UserAPIKeyAuth(key_alias="billing-app", user_id="u-1"),
|
||||
)
|
||||
|
||||
forwarded = mock_guardrail.apply_guardrail.await_args.kwargs["request_data"]["metadata"]
|
||||
assert forwarded == {
|
||||
"forbidden_topics": ["tax"],
|
||||
"user_api_key_alias": "billing-app",
|
||||
"user_api_key_user_id": "u-1",
|
||||
}
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_apply_guardrail_ignores_a_caller_supplied_user_api_key(mocker):
|
||||
"""The bare key field carries attribution as much as the prefixed ones, so a caller
|
||||
cannot set it either."""
|
||||
mock_guardrail = _patch_apply_guardrail_env(
|
||||
mocker,
|
||||
{"texts": ["ok"]},
|
||||
processed_data={
|
||||
"guardrail_name": "test-guardrail",
|
||||
"metadata": {"route": "/apply_guardrail", "user_api_key": "hash-of-the-real-key"},
|
||||
},
|
||||
)
|
||||
|
||||
request = ApplyGuardrailRequest(
|
||||
guardrail_name="test-guardrail",
|
||||
text="hello",
|
||||
metadata={"forbidden_topics": ["tax"], "user_api_key": "hash-of-someone-else"},
|
||||
)
|
||||
await apply_guardrail(
|
||||
fastapi_request=mocker.Mock(),
|
||||
request=request,
|
||||
user_api_key_dict=UserAPIKeyAuth(key_alias="billing-app", user_id="u-1"),
|
||||
)
|
||||
|
||||
forwarded = mock_guardrail.apply_guardrail.await_args.kwargs["request_data"]["metadata"]
|
||||
assert forwarded == {"forbidden_topics": ["tax"], "user_api_key": "hash-of-the-real-key"}
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_apply_guardrail_attaches_key_identity_when_caller_sends_no_metadata(mocker):
|
||||
"""A caller that sends no metadata still gets the authenticated identity
|
||||
forwarded, the same shape every LLM route gives a guardrail."""
|
||||
mock_guardrail = _patch_apply_guardrail_env(
|
||||
mocker,
|
||||
{"texts": ["ok"]},
|
||||
processed_data={"guardrail_name": "test-guardrail", "metadata": {"user_api_key_alias": "billing-app"}},
|
||||
)
|
||||
|
||||
request = ApplyGuardrailRequest(guardrail_name="test-guardrail", text="hello")
|
||||
await apply_guardrail(
|
||||
fastapi_request=mocker.Mock(),
|
||||
request=request,
|
||||
user_api_key_dict=UserAPIKeyAuth(),
|
||||
user_api_key_dict=UserAPIKeyAuth(key_alias="billing-app"),
|
||||
)
|
||||
|
||||
assert mock_guardrail.apply_guardrail.await_args.kwargs["request_data"] == {
|
||||
"metadata": {"user_api_key_alias": "billing-app"}
|
||||
}
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_apply_guardrail_sends_the_key_identity_when_the_caller_sends_no_metadata(mocker):
|
||||
"""A caller that sends no metadata still arrives with the key's identity attached, because
|
||||
pre-call processing fills it in. An empty request_data is not a shape the pipeline produces."""
|
||||
mock_guardrail = _patch_apply_guardrail_env(
|
||||
mocker,
|
||||
{"texts": ["ok"]},
|
||||
processed_data={"guardrail_name": "test-guardrail", "metadata": {"user_api_key_alias": "billing-app"}},
|
||||
)
|
||||
|
||||
request = ApplyGuardrailRequest(guardrail_name="test-guardrail", text="hello")
|
||||
await apply_guardrail(
|
||||
fastapi_request=mocker.Mock(),
|
||||
request=request,
|
||||
user_api_key_dict=UserAPIKeyAuth(key_alias="billing-app"),
|
||||
)
|
||||
|
||||
mock_guardrail.apply_guardrail.assert_awaited_once_with(
|
||||
inputs={"texts": ["hello"]},
|
||||
request_data={},
|
||||
request_data={"metadata": {"user_api_key_alias": "billing-app"}},
|
||||
input_type="request",
|
||||
)
|
||||
|
||||
|
|
|
|||
Loading…
Add table
Reference in a new issue