diff --git a/libs/prebuilt/langgraph/prebuilt/tool_node.py b/libs/prebuilt/langgraph/prebuilt/tool_node.py
index 00385f847..509584160 100644
--- a/libs/prebuilt/langgraph/prebuilt/tool_node.py
+++ b/libs/prebuilt/langgraph/prebuilt/tool_node.py
@@ -79,6 +79,7 @@ from langchain_core.tools.base import (
     TOOL_MESSAGE_BLOCK_TYPES,
     ToolException,
     _DirectlyInjectedToolArg,
+    _is_injected_arg_type,
     get_all_basemodel_annotations,
 )
 from langgraph._internal._runnable import RunnableCallable
@@ -611,6 +612,7 @@ class _InjectedArgs:
     state: dict[str, str | None]
     store: str | None
     runtime: str | None
+    all_injected_keys: set[str]
 
 
 class ToolNode(RunnableCallable):
@@ -1377,7 +1379,15 @@ class ToolNode(RunnableCallable):
         if injected.runtime:
             injected_args[injected.runtime] = tool_runtime
 
-        tool_call_copy["args"] = {**tool_call_copy["args"], **injected_args}
+        # Strip any caller-supplied values for injected args, then add
+        # back only trusted values. This prevents an LLM from forging
+        # hidden InjectedToolArg fields via ToolCall.args.
+        stripped_args = {
+            k: v
+            for k, v in tool_call_copy["args"].items()
+            if k not in injected.all_injected_keys
+        }
+        tool_call_copy["args"] = {**stripped_args, **injected_args}
         return tool_call_copy
 
     def _validate_tool_command(
@@ -1841,8 +1851,13 @@ def _get_all_injected_args(tool: BaseTool) -> _InjectedArgs:
     state_args: dict[str, str | None] = {}
     store_arg: str | None = None
     runtime_arg: str | None = None
+    all_injected_keys: set[str] = set()
 
     for name, type_ in all_annotations.items():
+        # Track all InjectedToolArg-annotated params (including custom subclasses)
+        if _is_injected_arg_type(type_):
+            all_injected_keys.add(name)
+
         # Check for runtime (special case: parameter named "runtime")
         if name == "runtime":
             runtime_arg = name
@@ -1866,4 +1881,5 @@ def _get_all_injected_args(tool: BaseTool) -> _InjectedArgs:
         state=state_args,
         store=store_arg,
         runtime=runtime_arg,
+        all_injected_keys=all_injected_keys,
     )
diff --git a/libs/prebuilt/tests/test_tool_node.py b/libs/prebuilt/tests/test_tool_node.py
index 506572ff3..72fd5bc7f 100644
--- a/libs/prebuilt/tests/test_tool_node.py
+++ b/libs/prebuilt/tests/test_tool_node.py
@@ -21,7 +21,7 @@ from langchain_core.messages import (
     ToolMessage,
 )
 from langchain_core.runnables.config import RunnableConfig
-from langchain_core.tools import BaseTool, ToolException
+from langchain_core.tools import BaseTool, InjectedToolArg, ToolException
 from langchain_core.tools import tool as dec_tool
 from langgraph.config import get_stream_writer
 from langgraph.errors import GraphBubbleUp, GraphInterrupt
@@ -2008,3 +2008,98 @@ async def test_tool_node_inject_runtime_dynamic_tool_via_wrap_tool_call_async()
     tool_message = result["messages"][-1]
     assert tool_message.content == "dynamic: x=42, tool_call_id=call_dynamic_2"
     assert tool_message.tool_call_id == "call_dynamic_2"
+
+
+# --- InjectedToolArg security tests ---
+
+
+def test_tool_node_strips_plain_injected_tool_arg() -> None:
+    """Plain InjectedToolArg values supplied by the LLM should be stripped."""
+
+    @dec_tool
+    def read_secret(
+        query: str,
+        auth: Annotated[dict, InjectedToolArg()],
+    ) -> str:
+        """Return secret data based on auth role."""
+        if auth.get("role") == "admin":
+            return "ADMIN_SECRET"
+        return "PUBLIC_DATA"
+
+    node = ToolNode([read_secret], handle_tool_errors=True)
+
+    # LLM tries to supply the hidden 'auth' field
+    tool_call = {
+        "name": "read_secret",
+        "args": {"query": "hello", "auth": {"role": "admin"}},
+        "id": "call-1",
+        "type": "tool_call",
+    }
+    msg = AIMessage("", tool_calls=[tool_call])
+    result = node.invoke({"messages": [msg]}, config=_create_config_with_runtime())
+    tool_message = result["messages"][-1]
+    # auth should have been stripped, so tool should fail (missing required arg)
+    assert "ADMIN_SECRET" not in tool_message.content
+
+
+def test_tool_node_strips_custom_injected_tool_arg_subclass() -> None:
+    """Custom InjectedToolArg subclasses should also be stripped."""
+
+    class InjectedAuth(InjectedToolArg):
+        pass
+
+    @dec_tool
+    def read_secret(
+        query: str,
+        auth: Annotated[dict, InjectedAuth()],
+    ) -> str:
+        """Return secret data based on auth role."""
+        if auth.get("role") == "admin":
+            return "ADMIN_SECRET"
+        return "PUBLIC_DATA"
+
+    node = ToolNode([read_secret], handle_tool_errors=True)
+
+    tool_call = {
+        "name": "read_secret",
+        "args": {"query": "hello", "auth": {"role": "admin"}},
+        "id": "call-1",
+        "type": "tool_call",
+    }
+    msg = AIMessage("", tool_calls=[tool_call])
+    result = node.invoke({"messages": [msg]}, config=_create_config_with_runtime())
+    tool_message = result["messages"][-1]
+    assert "ADMIN_SECRET" not in tool_message.content
+
+
+def test_tool_node_injected_state_overwrites_llm_value() -> None:
+    """InjectedState should use graph state, not LLM-supplied values."""
+
+    @dec_tool
+    def read_secret(
+        query: str,
+        auth: Annotated[dict, InjectedState("auth")],
+    ) -> str:
+        """Return secret data based on auth from graph state."""
+        if auth.get("role") == "admin":
+            return "ADMIN_SECRET"
+        return "PUBLIC_DATA"
+
+    node = ToolNode([read_secret])
+
+    # LLM tries to supply auth as admin
+    tool_call = {
+        "name": "read_secret",
+        "args": {"query": "hello", "auth": {"role": "admin"}},
+        "id": "call-1",
+        "type": "tool_call",
+    }
+    msg = AIMessage("", tool_calls=[tool_call])
+
+    # Graph state has auth as viewer
+    result = node.invoke(
+        {"messages": [msg], "auth": {"role": "viewer"}},
+        config=_create_config_with_runtime(),
+    )
+    tool_message = result["messages"][-1]
+    assert tool_message.content == "PUBLIC_DATA"