AIKIDO-2025-10510

TL;DR

Affected versions of this package are vulnerable to arbitrary code execution due to unsafe deserialization of user-controlled input when using plugins. Specifically, the model_revision parameter can be exploited to inject and execute arbitrary Python code, posing a significant security risk in environments where untrusted input reaches the pipeline function. The issue is addressed in the patched version by introducing the trust_remote_code parameter in the from_pretrained function, which controls whether remote code execution is allowed. By default, this parameter is set to False, preventing unauthorized or untrusted code from being executed.