diff --git a/tests/router_unit_tests/test_router_embedding_integration.py b/tests/router_unit_tests/test_router_embedding_integration.py
index e10ba0f0962..b295a351985 100644
--- a/tests/router_unit_tests/test_router_embedding_integration.py
+++ b/tests/router_unit_tests/test_router_embedding_integration.py
@@ -6,6 +6,7 @@ need to be properly propagated through the router to the LLM API.
 """
 
 import json
+from types import SimpleNamespace
 from unittest.mock import AsyncMock, MagicMock, patch
 
 import httpx
@@ -13,6 +14,7 @@ import pytest
 import respx
 
 import litellm
+import litellm.router_strategy.simple_shuffle as simple_shuffle
 from litellm import Router
 from litellm.llms.base_llm.vector_store.transformation import (
     LiteLLMVectorStoreEmbeddingExecutor,
@@ -383,6 +385,55 @@ class TestRouterEmbeddingIntegration:
             # The call should succeed
             mock_aembedding.assert_called_once()
 
+    def test_sync_embedding_respects_model_access_group_scoping(self, monkeypatch):
+        """
+        Regression test for #31260: sync Router._embedding must forward the
+        caller's request_kwargs to deployment selection. Without the forward,
+        _filter_deployments_by_model_access_groups short-circuits
+        (request_kwargs is None) and every deployment in the model group is
+        treated as globally accessible. random.choice is forced to the last
+        candidate so an unfiltered deployment list fails this test
+        deterministically instead of on 50% of runs.
+        """
+        model_list = [
+            {
+                "model_name": "grouped-embed",
+                "litellm_params": {
+                    "model": "text-embedding-3-small",
+                    "api_key": "eng-key",
+                },
+                "model_info": {"access_groups": ["engineering"]},
+            },
+            {
+                "model_name": "grouped-embed",
+                "litellm_params": {
+                    "model": "text-embedding-3-small",
+                    "api_key": "fin-key",
+                },
+                "model_info": {"access_groups": ["finance"]},
+            },
+        ]
+
+        router = Router(model_list=model_list)
+        engineering_auth = SimpleNamespace(models=["engineering"], team_models=[])
+
+        fake_random = MagicMock()
+        fake_random.choice.side_effect = lambda seq: seq[-1]
+
+        monkeypatch.setattr(litellm, "embedding", MagicMock(return_value=MagicMock(data=[{"embedding": [0.1, 0.2]}])))
+        monkeypatch.setattr(simple_shuffle, "random", fake_random)
+
+        mock_embedding = litellm.embedding
+        router.embedding(
+            model="grouped-embed",
+            input=["hello"],
+            metadata={"user_api_key_auth": engineering_auth},
+        )
+
+        # The finance deployment must have been filtered out; only the
+        # engineering deployment can be selected.
+        assert mock_embedding.call_args[1]["api_key"] == "eng-key"
+
     def test_embedding_with_timeout_from_router(self):
         """
         Test that timeout settings from router config are propagated.
diff --git a/tests/test_litellm/caching/test_caching.py b/tests/test_litellm/caching/test_caching.py
index 4d0ec0fb677..7fb2f9a75f2 100644
--- a/tests/test_litellm/caching/test_caching.py
+++ b/tests/test_litellm/caching/test_caching.py
@@ -1,5 +1,6 @@
 import logging
 import re
+from unittest.mock import MagicMock
 
 import pytest
 
@@ -252,3 +253,33 @@ def test_exact_cache_key_includes_anthropic_messages_params(anthropic_param):
     assert baseline != cache.get_cache_key(
         model="claude-sonnet-4-5", messages=messages, **anthropic_param
     )
+
+
+def test_sync_get_cache_forwards_metadata_subset_to_backend():
+    """
+    Regression test for #31260: the sync cache read path used to flatten
+    metadata to {}, so semantic-cache backends never saw the caller's auth
+    context (team id / user_api_key_auth) on a get, while the write path and
+    the async read path forwarded full metadata. The metadata subset must be
+    preserved so sync semantic-cache lookups can scope and authenticate the
+    same way async ones do.
+    """
+    cache = Cache(type=LiteLLMCacheType.LOCAL)
+    backend = MagicMock()
+    backend.get_cache.return_value = None
+
+    cache.get_cache(
+        dynamic_cache_object=backend,
+        cache={"use-cache": True},
+        model="text-embedding-3-small",
+        input=["hello"],
+        metadata={
+            "user_api_key_team_id": "team-a",
+            "user_api_key": "k-abc",
+        },
+    )
+
+    assert backend.get_cache.called
+    sent_metadata = backend.get_cache.call_args.kwargs["metadata"]
+    assert sent_metadata["user_api_key_team_id"] == "team-a"
+    assert sent_metadata["user_api_key"] == "k-abc"