Merge facbc8734c into e9b1c11098

litellm/proxy/auth/auth_checks.py

@@ -5517,6 +5517,29 @@ async def stamp_matched_model_access_groups(
     return matched
 
 
+def _log_key_model_access_denied(
+    valid_token: UserAPIKeyAuth,
+    model: str | Sequence[str],
+    allowed_models: Sequence[object] | None,
+) -> None:
+    """Log which key/user was denied which model, server-side only.
+
+    Key identity (alias / key_name / user_id) is PII: it goes to the server
+    warning log only — the client-facing exception message stays generic. The
+    speculative inner model checks (``_can_object_call_model``, which run before
+    the access-group fallback) deliberately log nothing, so a request that is
+    finally authorized via fallback is not miscounted as a denial.
+    """
+    verbose_proxy_logger.warning(
+        "Model access denied: key_alias=%s key_name=%s user_id=%s requested_model=%s allowed_models=%s",
+        valid_token.key_alias,
+        valid_token.key_name,
+        valid_token.user_id,
+        model,
+        list(allowed_models or []),
+    )
+
+
 async def can_key_call_model(
     model: str | list[str],
     llm_model_list: Sequence[object] | None,
@@ -5556,15 +5579,20 @@ async def can_key_call_model(
                 prisma_client=prisma_client,
             )
             if models_from_groups:
-                return can_object_call_model(
-                    model=model,
-                    llm_router=llm_router,
-                    models=models_from_groups,
-                    team_model_aliases=valid_token.team_model_aliases,
-                    team_id=valid_token.team_id,
-                    key_model_aliases=key_model_aliases_for_auth_check(valid_token),
-                    object_type="key",
-                )
+                try:
+                    return can_object_call_model(
+                        model=model,
+                        llm_router=llm_router,
+                        models=models_from_groups,
+                        team_model_aliases=valid_token.team_model_aliases,
+                        team_id=valid_token.team_id,
+                        key_model_aliases=key_model_aliases_for_auth_check(valid_token),
+                        object_type="key",
+                    )
+                except ProxyException:
+                    _log_key_model_access_denied(valid_token, model, models_from_groups)
+                    raise
+        _log_key_model_access_denied(valid_token, model, key_models)
         raise
 
 

tests/unit/proxy/auth/test_key_model_access_denial_logging.py

@@ -0,0 +1,120 @@
+"""Regression tests: when a key is denied model access, the proxy must log which
+key/user was denied which model (server-side warning) — while the client-facing
+exception stays free of key identity, and successful access-group fallbacks are
+not miscounted as denials."""
+
+import logging
+from unittest.mock import AsyncMock, patch
+
+import pytest
+from litellm.proxy._types import UserAPIKeyAuth
+from litellm.proxy.auth.auth_checks import can_key_call_model
+
+_LLM_MODEL_LIST = [
+    {
+        "model_name": "openai/*",
+        "litellm_params": {"model": "openai/*", "api_key": "test-api-key"},
+        "model_info": {
+            "id": "e6e7006f83029df40ebc02ddd068890253f4cd3092bcb203d3d8e6f6f606f30f",
+            "db_model": False,
+            "access_groups": ["public-openai-models"],
+        },
+    },
+    {
+        "model_name": "openai/gpt-4o",
+        "litellm_params": {"model": "openai/gpt-4o", "api_key": "test-api-key"},
+        "model_info": {
+            "id": "0cfcd87f2cb12a783a466888d05c6c89df66db23e01cecd75ec0b83aed73c9ad",
+            "db_model": False,
+            "access_groups": ["private-openai-models"],
+        },
+    },
+]
+
+
+def _auth(**overrides) -> UserAPIKeyAuth:
+    kwargs = {
+        "models": ["public-openai-models"],
+        "key_alias": "prod-key",
+        "key_name": "prod",
+        "user_id": "user-123",
+    }
+    kwargs.update(overrides)
+    return UserAPIKeyAuth(**kwargs)
+
+
+def _denial_records(caplog) -> list:
+    return [r for r in caplog.records if "Model access denied" in r.getMessage()]
+
+
+@pytest.mark.asyncio
+async def test_denial_without_fallback_logs_key_context(caplog):
+    """A key denied natively with no usable access-group fallback must log the
+    key identity + requested/allowed models server-side."""
+    import litellm
+
+    router = litellm.Router(model_list=_LLM_MODEL_LIST)
+    caplog.set_level(logging.WARNING, logger="LiteLLM Proxy")
+
+    with pytest.raises(Exception, match="is not available for this API key") as exc_info:
+        await can_key_call_model(
+            model="openai/gpt-4o",
+            llm_model_list=_LLM_MODEL_LIST,
+            valid_token=_auth(),
+            llm_router=router,
+        )
+
+    records = _denial_records(caplog)
+    assert records, "denial must produce a server-side Model access denied warning"
+    message = records[-1].getMessage()
+    assert "key_alias=prod-key" in message
+    assert "key_name=prod" in message
+    assert "user_id=user-123" in message
+    assert "requested_model=openai/gpt-4o" in message
+    assert "public-openai-models" in message
+    assert "prod-key" not in str(exc_info.value), "client-facing message must stay free of key identity"
+
+
+@pytest.mark.asyncio
+async def test_denial_after_access_group_fallback_logs_key_context(caplog):
+    """When the access-group fallback is attempted and also denies, the denial
+    must still be logged exactly once."""
+    import litellm
+
+    router = litellm.Router(model_list=_LLM_MODEL_LIST)
+    caplog.set_level(logging.WARNING, logger="LiteLLM Proxy")
+
+    with patch(
+        "litellm.proxy.auth.auth_checks.get_models_from_access_groups",
+        new_callable=AsyncMock,
+        return_value=["unrelated-model-group"],
+    ):
+        with pytest.raises(Exception, match="is not available for this API key"):
+            await can_key_call_model(
+                model="openai/gpt-4o",
+                llm_model_list=_LLM_MODEL_LIST,
+                valid_token=_auth(access_group_ids=["ag-1"]),
+                llm_router=router,
+            )
+
+    records = _denial_records(caplog)
+    assert len(records) == 1, "fallback denial path must log exactly one denial"
+    assert "key_alias=prod-key" in records[0].getMessage()
+    assert "unrelated-model-group" in records[0].getMessage()
+
+
+@pytest.mark.asyncio
+async def test_authorized_model_logs_no_denial(caplog):
+    import litellm
+
+    router = litellm.Router(model_list=_LLM_MODEL_LIST)
+    caplog.set_level(logging.WARNING, logger="LiteLLM Proxy")
+
+    await can_key_call_model(
+        model="openai/gpt-4o-mini",
+        llm_model_list=_LLM_MODEL_LIST,
+        valid_token=_auth(models=["*"]),
+        llm_router=router,
+    )
+
+    assert not _denial_records(caplog), "authorized request must not log a denial"

tests/unit/proxy/auth/test_user_api_key_auth.py

@@ -548,8 +548,6 @@ def test_get_api_key_from_custom_header_different_casing():
     )
 
 
-
-
 @pytest.mark.parametrize(
     "user_role, auth_user_id, requested_user_id, expected_result",
     [
@@ -2069,7 +2067,10 @@ async def test_user_api_key_auth_websocket_logs_a_model_access_denial_as_one_war
     proxy_records: Final = [record for record in caplog.records if record.name == verbose_proxy_logger.name]
     assert [record.getMessage() for record in proxy_records if record.exc_info is not None] == []
     assert [record.getMessage() for record in proxy_records if record.levelno == logging.WARNING] == [
-        "key not allowed to access model. This key can only access models=['gpt-5.4-mini']. Tried to access gpt-5.4"
+        # Server-side denial context from can_key_call_model (key identity stays out of the
+        # client-facing exception), followed by the request-path denial warning.
+        "Model access denied: key_alias=None key_name=None user_id=None requested_model=gpt-5.4 allowed_models=['gpt-5.4-mini']",
+        "key not allowed to access model. This key can only access models=['gpt-5.4-mini']. Tried to access gpt-5.4",
     ]