diff --git a/litellm/types/proxy/guardrails/guardrail_hooks/generic_guardrail_api.py b/litellm/types/proxy/guardrails/guardrail_hooks/generic_guardrail_api.py
index 7f167b8b660..b3f6dae40eb 100644
--- a/litellm/types/proxy/guardrails/guardrail_hooks/generic_guardrail_api.py
+++ b/litellm/types/proxy/guardrails/guardrail_hooks/generic_guardrail_api.py
@@ -40,21 +40,25 @@ class GenericGuardrailAPIOptionalParams(BaseModel):
     )
 
     streaming_end_of_stream_only: Optional[bool] = Field(
-        default=False,
+        default=None,
         description=(
-            "If False (default), the guardrail runs on sampled chunks during the stream "
-            "at the cadence set by streaming_sampling_rate, and an in-flight BLOCKED "
-            "stops further chunks from streaming. If True, the guardrail runs once at "
-            "end of stream over the assembled response; lower cost and latency, but "
-            "flagged content has already streamed to the client before the terminal block."
+            "If False (default when unset), the guardrail runs on sampled chunks during "
+            "the stream at the cadence set by streaming_sampling_rate, and an in-flight "
+            "BLOCKED stops further chunks from streaming. If True, the guardrail runs "
+            "once at end of stream over the assembled response; lower cost and latency, "
+            "but flagged content has already streamed to the client before the terminal "
+            "block. Defaults are applied in GenericGuardrailAPI.__init__ when None so "
+            "unset optional_params does not shadow top-level litellm_params."
         ),
     )
 
     streaming_sampling_rate: Optional[int] = Field(
-        default=5,
+        default=None,
         description=(
             "When streaming_end_of_stream_only is False, the guardrail runs every Nth "
-            "streamed chunk. Ignored when streaming_end_of_stream_only is True."
+            "streamed chunk. Ignored when streaming_end_of_stream_only is True. "
+            "Defaults to 5 in GenericGuardrailAPI.__init__ when None so unset "
+            "optional_params does not shadow top-level litellm_params."
         ),
     )
 
diff --git a/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_generic_guardrail_api.py b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_generic_guardrail_api.py
index d3263d71697..ad5f1f3e30d 100644
--- a/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_generic_guardrail_api.py
+++ b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_generic_guardrail_api.py
@@ -1124,6 +1124,73 @@ class TestGenericGuardrailAPIStreamingConfig:
         assert guardrail.streaming_end_of_stream_only is False
         assert guardrail.streaming_sampling_rate == 3
 
+    def test_initialize_guardrail_optional_params_defaults_do_not_shadow_top_level(
+        self,
+    ):
+        """Top-level streaming knobs win when optional_params only carries siblings."""
+        from litellm.proxy.guardrails.guardrail_hooks.generic_guardrail_api import (
+            initialize_guardrail,
+        )
+        from litellm.types.guardrails import LitellmParams
+        from litellm.types.proxy.guardrails.guardrail_hooks.generic_guardrail_api import (
+            GenericGuardrailAPIOptionalParams,
+        )
+
+        litellm_params = LitellmParams(
+            guardrail="generic_guardrail_api",
+            mode="post_call",
+            api_base="https://api.test.guardrail.com",
+            default_on=False,
+        )
+        litellm_params.streaming_end_of_stream_only = True  # type: ignore[attr-defined]
+        litellm_params.streaming_sampling_rate = 2  # type: ignore[attr-defined]
+        # Sibling optional_params only; streaming fields stay at Pydantic default None.
+        litellm_params.optional_params = GenericGuardrailAPIOptionalParams(  # type: ignore[attr-defined]
+            additional_provider_specific_params={"tenant": "acme"},
+        )
+
+        guardrail_config = {"guardrail_name": "test-generic-streaming-mixed"}
+
+        with patch(
+            "litellm.logging_callback_manager.add_litellm_callback"
+        ):
+            guardrail = initialize_guardrail(litellm_params, guardrail_config)
+
+        assert guardrail.streaming_end_of_stream_only is True
+        assert guardrail.streaming_sampling_rate == 2
+
+    def test_initialize_guardrail_explicit_optional_params_streaming_wins(self):
+        from litellm.proxy.guardrails.guardrail_hooks.generic_guardrail_api import (
+            initialize_guardrail,
+        )
+        from litellm.types.guardrails import LitellmParams
+        from litellm.types.proxy.guardrails.guardrail_hooks.generic_guardrail_api import (
+            GenericGuardrailAPIOptionalParams,
+        )
+
+        litellm_params = LitellmParams(
+            guardrail="generic_guardrail_api",
+            mode="post_call",
+            api_base="https://api.test.guardrail.com",
+            default_on=False,
+        )
+        litellm_params.streaming_end_of_stream_only = False  # type: ignore[attr-defined]
+        litellm_params.streaming_sampling_rate = 9  # type: ignore[attr-defined]
+        litellm_params.optional_params = GenericGuardrailAPIOptionalParams(  # type: ignore[attr-defined]
+            streaming_end_of_stream_only=True,
+            streaming_sampling_rate=1,
+        )
+
+        guardrail_config = {"guardrail_name": "test-generic-streaming-nested-wins"}
+
+        with patch(
+            "litellm.logging_callback_manager.add_litellm_callback"
+        ):
+            guardrail = initialize_guardrail(litellm_params, guardrail_config)
+
+        assert guardrail.streaming_end_of_stream_only is True
+        assert guardrail.streaming_sampling_rate == 1
+
 
 class TestGenericGuardrailAPIStreamingViaUnified:
     """Streaming output checks routed through UnifiedLLMGuardrails."""