diff --git a/litellm/proxy/auth/auth_checks.py b/litellm/proxy/auth/auth_checks.py
index 9dce42bd3a9..e4e2ceb6621 100644
--- a/litellm/proxy/auth/auth_checks.py
+++ b/litellm/proxy/auth/auth_checks.py
@@ -2697,7 +2697,11 @@ def compute_effective_models(
 
     effective = union(team_defaults, member_models), capped by team_pool.
     - If neither defaults nor overrides are set, falls back to team_pool (backward compat).
-    - If cap empties the list (all stale), falls back to team_pool (NOT [] which = allow-all).
+    - If cap empties the list (all overrides/defaults are stale), falls back to team_pool
+      rather than returning [] (which would mean "allow all"). This is a deliberate security
+      trade-off: a member with entirely stale overrides gets team-pool access (the team's
+      restriction is still enforced) instead of unrestricted access. Admins should clean up
+      stale overrides via /team/member_update when narrowing team.models.
     - team_pool=[] means "allow all" — cap is skipped.
     """
     # dict.fromkeys preserves insertion order while deduplicating
diff --git a/litellm/proxy/management_endpoints/team_endpoints.py b/litellm/proxy/management_endpoints/team_endpoints.py
index a1cf563144e..5b47861899f 100644
--- a/litellm/proxy/management_endpoints/team_endpoints.py
+++ b/litellm/proxy/management_endpoints/team_endpoints.py
@@ -2533,7 +2533,7 @@ async def team_member_update(  # noqa: PLR0915
     from litellm.proxy.proxy_server import user_api_key_cache
 
     _cache_key = f"team_membership:{received_user_id}:{data.team_id}"
-    user_api_key_cache.delete_cache(key=_cache_key)
+    await user_api_key_cache.async_delete_cache(key=_cache_key)
 
     ### update team member role
     # Resolve the effective models for this member (from the authoritative
diff --git a/litellm/proxy/management_helpers/utils.py b/litellm/proxy/management_helpers/utils.py
index 15385b2a42b..afb0c59bc2c 100644
--- a/litellm/proxy/management_helpers/utils.py
+++ b/litellm/proxy/management_helpers/utils.py
@@ -140,7 +140,7 @@ async def handle_budget_for_entity(
         return existing_budget_id
 
 
-async def add_new_member(
+async def add_new_member(  # noqa: PLR0915
     new_member: Member,
     max_budget_in_team: Optional[float],
     prisma_client: PrismaClient,
@@ -271,6 +271,13 @@ async def add_new_member(
             **_returned_team_membership.model_dump()
         )
 
+        # Invalidate any stale cached membership for this user+team pair
+        # (e.g., from a previous add/remove cycle).
+        from litellm.proxy.proxy_server import user_api_key_cache
+
+        _cache_key = f"team_membership:{returned_user.user_id}:{team_id}"
+        await user_api_key_cache.async_delete_cache(key=_cache_key)
+
     if returned_user is None:
         raise Exception("Unable to update user table with membership information!")