diff --git a/litellm/proxy/litellm_pre_call_utils.py b/litellm/proxy/litellm_pre_call_utils.py
index 1d3ef2e10c2..064538ef3bf 100644
--- a/litellm/proxy/litellm_pre_call_utils.py
+++ b/litellm/proxy/litellm_pre_call_utils.py
@@ -1500,10 +1500,13 @@ def add_guardrails_from_policy_engine(
     Add guardrails from the policy engine based on request context.
 
     This function:
-    1. Gets matching policies based on team_alias, key_alias, and model
-    2. Resolves guardrails from matching policies (including inheritance)
-    3. Adds guardrails to request metadata
-    4. Tracks applied policies in metadata for response headers
+    1. Extracts "policies" from request body (if present) for dynamic policy application
+    2. Gets matching policies based on team_alias, key_alias, and model (via attachments)
+    3. Combines dynamic policies with attachment-based policies
+    4. Resolves guardrails from all policies (including inheritance)
+    5. Adds guardrails to request metadata
+    6. Tracks applied policies in metadata for response headers
+    7. Removes "policies" from request body so it's not forwarded to LLM provider
 
     Args:
         data: The request data to update
@@ -1519,6 +1522,10 @@ def add_guardrails_from_policy_engine(
     from litellm.proxy.policy_engine.policy_resolver import PolicyResolver
     from litellm.types.proxy.policy_engine import PolicyMatchContext
 
+    # Extract dynamic policies from request body (if present)
+    # These will be combined with attachment-based policies
+    request_body_policies = data.pop("policies", None)
+
     registry = get_policy_registry()
     verbose_proxy_logger.debug(
         f"Policy engine: registry initialized={registry.is_initialized()}, "
@@ -1545,12 +1552,18 @@ def add_guardrails_from_policy_engine(
 
     verbose_proxy_logger.debug(f"Policy engine: matched policies via attachments: {matching_policy_names}")
 
-    if not matching_policy_names:
+    # Combine attachment-based policies with dynamic request body policies
+    all_policy_names = set(matching_policy_names)
+    if request_body_policies and isinstance(request_body_policies, list):
+        all_policy_names.update(request_body_policies)
+        verbose_proxy_logger.debug(f"Policy engine: added dynamic policies from request body: {request_body_policies}")
+
+    if not all_policy_names:
         return
 
     # Filter to only policies whose conditions match the context
     applied_policy_names = PolicyMatcher.get_policies_with_matching_conditions(
-        policy_names=matching_policy_names,
+        policy_names=list(all_policy_names),
         context=context,
     )
 
diff --git a/tests/test_litellm/proxy/test_litellm_pre_call_utils.py b/tests/test_litellm/proxy/test_litellm_pre_call_utils.py
index b9485a2e4cb..da6a5aeab09 100644
--- a/tests/test_litellm/proxy/test_litellm_pre_call_utils.py
+++ b/tests/test_litellm/proxy/test_litellm_pre_call_utils.py
@@ -1548,3 +1548,50 @@ def test_add_guardrails_from_policy_engine():
     policy_registry._initialized = False
     attachment_registry._attachments = []
     attachment_registry._initialized = False
+
+
+def test_add_guardrails_from_policy_engine_accepts_dynamic_policies_and_pops_from_data():
+    """
+    Test that add_guardrails_from_policy_engine accepts dynamic 'policies' from the request body
+    and removes them to prevent forwarding to the LLM provider.
+    
+    This is critical because 'policies' is a LiteLLM proxy-specific parameter that should
+    not be sent to the actual LLM API (e.g., OpenAI, Anthropic, etc.).
+    """
+    from litellm.proxy.policy_engine.policy_registry import get_policy_registry
+
+    # Setup test data with 'policies' in the request body
+    data = {
+        "model": "gpt-4",
+        "messages": [{"role": "user", "content": "Hello"}],
+        "policies": ["PII-POLICY-GLOBAL", "HIPAA-POLICY"],  # Dynamic policies - should be accepted and removed
+        "metadata": {},
+    }
+
+    user_api_key_dict = UserAPIKeyAuth(
+        api_key="test-key",
+        team_alias="test-team",
+        key_alias="test-key",
+    )
+
+    # Initialize empty policy registry (we're just testing the accept and pop behavior)
+    policy_registry = get_policy_registry()
+    policy_registry._policies = {}
+    policy_registry._initialized = False
+
+    # Call the function - should accept dynamic policies and not raise an error
+    add_guardrails_from_policy_engine(
+        data=data,
+        metadata_variable_name="metadata",
+        user_api_key_dict=user_api_key_dict,
+    )
+
+    # Verify that 'policies' was removed from the request body
+    assert "policies" not in data, "'policies' should be removed from request body to prevent forwarding to LLM provider"
+
+    # Verify that other fields are preserved
+    assert "model" in data
+    assert data["model"] == "gpt-4"
+    assert "messages" in data
+    assert data["messages"] == [{"role": "user", "content": "Hello"}]
+    assert "metadata" in data