diff --git a/src/google/adk/tools/bigtable/bigtable_toolset.py b/src/google/adk/tools/bigtable/bigtable_toolset.py
index eaa1dae61d8..0aef34d288f 100644
--- a/src/google/adk/tools/bigtable/bigtable_toolset.py
+++ b/src/google/adk/tools/bigtable/bigtable_toolset.py
@@ -54,6 +54,12 @@ class BigtableParameterizedViewTool(GoogleTool):
       pass it as `view_parameters={"user_id": user_id}`.
       This securely restricts query execution to the logged-in user's data
       without exposing the `user_id` parameter to the LLM.
+
+      Only names that resolve to an attribute of tool_context itself (such as
+      user_id) are honored. tool_context.state is not consulted: state is
+      writable by the caller, so a name that fell back to state could be
+      overridden to run the query in a different user's scope, defeating the
+      view's own row-level restriction.
   """
 
   def __init__(
@@ -71,8 +77,9 @@ def __init__(
         credentials_config: The credentials configuration.
         tool_settings: The tool settings.
         view_parameter_names: A list of parameter names to resolve from
-          tool_context and pass into view_parameters. This is configured on the
-          toolset (BigtableToolset) and forwarded here.
+          tool_context's own attributes (not tool_context.state) and pass
+          into view_parameters. This is configured on the toolset
+          (BigtableToolset) and forwarded here.
     """
     super().__init__(
         func=func,
@@ -101,12 +108,14 @@ async def _run_async_with_credential(
     if "_view_parameters" in signature.parameters and self.view_parameter_names:
       view_params = {}
       for param_name in self.view_parameter_names:
-        # 1. Check if it's a strongly-typed top-level property (like 'user_id')
+        # Only resolve from strongly-typed, framework-set attributes of
+        # tool_context (like user_id). tool_context.state is deliberately not
+        # consulted here: it is writable by the caller, and a view parameter
+        # taken from there would let the caller pick which user's (or
+        # tenant's) rows the parameterized view returns, defeating the
+        # view's own row-level restriction.
         if (val := getattr(tool_context, param_name, None)) is not None:
           view_params[param_name] = val
-        # 2. Fallback to checking application-level session state
-        elif tool_context.state and param_name in tool_context.state:
-          view_params[param_name] = tool_context.state[param_name]
 
       args_to_call["_view_parameters"] = view_params
     return await super()._run_async_with_credential(
diff --git a/tests/unittests/tools/bigtable/test_bigtable_toolset.py b/tests/unittests/tools/bigtable/test_bigtable_toolset.py
index aa5fd440be2..ccdbc849b2d 100644
--- a/tests/unittests/tools/bigtable/test_bigtable_toolset.py
+++ b/tests/unittests/tools/bigtable/test_bigtable_toolset.py
@@ -253,18 +253,27 @@ def mock_execute_sql(_view_parameters=None):
 
 
 @pytest.mark.asyncio
-async def test_bigtable_parameterized_view_tool_execution_session_state_fallback():
-  """Test that BigtableParameterizedViewTool falls back to tool_context.state for custom parameters."""
+async def test_bigtable_parameterized_view_tool_ignores_state_for_custom_parameters():
+  """Test that BigtableParameterizedViewTool never resolves view parameters
+
+  from tool_context.state, even when a name matching view_parameter_names is
+  present there. tool_context.state is writable by the caller, so a view
+  parameter taken from it (e.g. a caller-chosen tenant_id) would let the
+  caller pick whose rows a parameterized view returns, defeating the view's
+  own row-level restriction. See the class docstring for the security
+  rationale.
+  """
 
   def mock_execute_sql(_view_parameters=None):
     return {"status": "SUCCESS", "_view_parameters": _view_parameters}
 
-  # Create session with application-level state
+  # A caller-controlled tenant_id sitting in session state, as if a prior
+  # tool call (or the model itself) had written it there.
   session = Session(
       id="session-1",
       app_name="test-app",
       user_id="user-123",
-      state={"tenant_id": "tenant-xyz"},
+      state={"tenant_id": "attacker-chosen-tenant"},
   )
 
   invocation_context = mock.create_autospec(InvocationContext, instance=True)
@@ -273,7 +282,8 @@ def mock_execute_sql(_view_parameters=None):
 
   tool_context = Context(invocation_context=invocation_context)
 
-  # Ensure 'tenant_id' is NOT a top-level property or attribute on tool_context
+  # Confirm the test setup: 'tenant_id' is not a top-level attribute of
+  # tool_context, only a key in its (caller-writable) state.
   assert not hasattr(tool_context, "tenant_id")
   assert "tenant_id" in tool_context.state
 
@@ -290,9 +300,10 @@ def mock_execute_sql(_view_parameters=None):
       tool_context=tool_context,
   )
 
+  # tenant_id is silently omitted rather than taken from state.
   assert res == {
       "status": "SUCCESS",
-      "_view_parameters": {"tenant_id": "tenant-xyz"},
+      "_view_parameters": {},
   }
 
 
@@ -330,11 +341,12 @@ def mock_execute_sql(_view_parameters=None):
       tool_context=tool_context,
   )
 
+  # Only user_id resolves, since it is the only one of the three that is a
+  # real tool_context attribute; tenant_id and agent_id are only present in
+  # state, which is never consulted, so they are silently omitted.
   assert res == {
       "status": "SUCCESS",
       "_view_parameters": {
           "user_id": "user-123",
-          "tenant_id": "tenant-xyz",
-          "agent_id": "agent-123",
       },
   }