The agent engineering platform.
Python agentsaiai-agentsanthropicchatgptdeepagentsenterpriseframeworkgeminigenerative-ailangchainlanggraphllmmultiagentopen-sourceopenaipydanticpythonragtypescript
docs(core): update load note to be precise (#35309)
Eugene Yurtsev8323f556d9c6
diff --git a/libs/core/langchain_core/load/load.py b/libs/core/langchain_core/load/load.pyindex eb4207f1f9..12bbf7a409 100644--- a/libs/core/langchain_core/load/load.py+++ b/libs/core/langchain_core/load/load.py@@ -13,13 +13,19 @@ allowlist. If the class is not in the allowlist, deserialization raises a `Value## Security model-!!! warning "Do not use with untrusted input"-- These functions instantiate Python objects and can trigger side effects- such as network calls, file operations, or environment variable access- during deserialization. **Never call `load()` or `loads()` on data from an- untrusted or unauthenticated source.** Even with the allowlist, allowed- classes may perform dangerous operations during `__init__`.+!!! warning "Exercise caution with untrusted input"++ These functions deserialize by instantiating Python objects, which means+ constructors (`__init__`) and validators may run and can trigger side effects.+ With the default settings, deserialization is restricted to a core allowlist+ of `langchain_core` types (for example: messages, documents, and prompts)+ defined in `langchain_core.load.mapping`.++ If you broaden `allowed_objects` (for example, by using `'all'` or adding+ additional classes), treat the serialized payload as a manifest and only+ deserialize data that comes from a trusted source. A crafted payload that+ is allowed to instantiate unintended classes could cause network calls,+ file operations, or environment variable access during `__init__`.The `allowed_objects` parameter controls which classes can be deserialized: