Skip to content

The agent engineering platform.

fix(core): correct misleading jinja2 sandboxing comment (#35183)

Yi Liu6a07b3acdadf
diff --git a/libs/core/langchain_core/prompts/string.py b/libs/core/langchain_core/prompts/string.py
index a7e1263544..0e3a29d75e 100644
--- a/libs/core/langchain_core/prompts/string.py
+++ b/libs/core/langchain_core/prompts/string.py
@@ -64,9 +64,11 @@ def jinja2_formatter(template: str, /, **kwargs: Any) -> str:
)
raise ImportError(msg)
- # Use a restricted sandbox that blocks ALL attribute/method access
- # Only simple variable lookups like {{variable}} are allowed
- # Attribute access like {{variable.attr}} or {{variable.method()}} is blocked
+ # Use Jinja2's SandboxedEnvironment which blocks access to dunder attributes
+ # (e.g., __class__, __globals__) to prevent sandbox escapes.
+ # Note: regular attribute access (e.g., {{obj.attr}}) and method calls are
+ # still allowed. This is a best-effort measure — do not use with untrusted
+ # templates.
return SandboxedEnvironment().from_string(template).render(**kwargs)