From 6efe96bea6ea720a9e31c76411e713d1afada59b Mon Sep 17 00:00:00 2001
From: Nick Hollon <nick.hollon@langchain.dev>
Date: Tue, 5 May 2026 16:29:40 -0400
Subject: [PATCH] fix(langchain): restrict deserialization in
 `langchain.storage._lc_store` (#37209)

---
 libs/langchain/langchain/storage/_lc_store.py | 35 +++++++++++++++++--
 1 file changed, 33 insertions(+), 2 deletions(-)

diff --git a/libs/langchain/langchain/storage/_lc_store.py b/libs/langchain/langchain/storage/_lc_store.py
index 24603b93599..9b807a8841f 100644
--- a/libs/langchain/langchain/storage/_lc_store.py
+++ b/libs/langchain/langchain/storage/_lc_store.py
@@ -24,7 +24,7 @@ def _dump_document_as_bytes(obj: Any) -> bytes:
 
 def _load_document_from_bytes(serialized: bytes) -> Document:
     """Return a document from a bytes representation."""
-    obj = loads(serialized.decode("utf-8"))
+    obj = loads(serialized.decode("utf-8"), allowed_objects=[Document])
     if not isinstance(obj, Document):
         msg = f"Expected a Document instance. Got {type(obj)}"
         raise TypeError(msg)
@@ -32,7 +32,14 @@ def _load_document_from_bytes(serialized: bytes) -> Document:
 
 
 def _load_from_bytes(serialized: bytes) -> Serializable:
-    """Return a document from a bytes representation."""
+    """Return a ``Serializable`` from a bytes representation."""
+    # The default allowlist (``'core'``) is unsafe with untrusted input - a
+    # tampered byte payload can reconstruct any core class with
+    # attacker-controlled kwargs (custom ``base_url``, headers, model name,
+    # etc.). The byte store backing this loader must be treated as a trust
+    # boundary - see the danger note on ``create_lc_store``. If the store
+    # can be written to by anyone you do not already trust, use
+    # ``create_kv_docstore`` instead.
     return loads(serialized.decode("utf-8"))
 
 
@@ -51,6 +58,30 @@ def create_lc_store(
 ) -> BaseStore[str, Serializable]:
     """Create a store for langchain serializable objects from a bytes store.
 
+    .. danger::
+
+        Treat the underlying byte store as a trust boundary.
+
+        Reads from this store are deserialized with
+        ``langchain_core.load.loads``, which instantiates Python objects
+        from the stored payload. The same threat model applies: a payload
+        can carry constructor kwargs (custom ``base_url``, headers, model
+        name, etc.) that get applied during ``__init__``, so the bytes are
+        effectively executable configuration rather than plain data.
+
+        **Never back this store with anything an attacker can write to** -
+        for example a shared cache that other tenants can populate, an
+        S3 bucket without strict write controls, or a Redis instance
+        reused across trust boundaries. A single tampered value will
+        instantiate attacker-controlled classes the next time the store
+        is read.
+
+        If you cannot guarantee the store is write-restricted to your own
+        process, use ``create_kv_docstore`` instead - it pins
+        ``allowed_objects=[Document]`` so a tampered value can at worst
+        produce a ``Document``, never a chat model or LLM with a
+        redirected endpoint.
+
     Args:
         store: A bytes store to use as the underlying store.
         key_encoder: A function to encode keys; if None uses identity function.