GHSA-86cj-95qr-2p4f: picklescan: detection bypass enables PyTorch model RCE

### Summary Using torch._dynamo.guards.GuardBuilder.get function, which is a pytorch 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 torch._dynamo.guards.GuardBuilder.get 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 ``` import types import torch._dynamo.guards as guards class EvilTorchDynamoGuardsGet: def __reduce__(self): fake_self = types.SimpleNamespace(scope={}) name = "__import__('os').system('whoami')" return guards.GuardBuilder.get, (fake_self, name) ``` ### 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 targets an MLOps team that accepts community-contributed models. They publish a seemingly legitimate PyTorch model (e.g., a fine-tuned LLM adapter) to Hugging Face or a private registry. The model file contains a crafted `__reduce__` that invokes `torch._dynamo.guards.GuardBuilder.get` with an embedded shell command. The victim's intake pipeline runs picklescan on the file, receives a clean result, and promotes the model to staging. On first inference load, the malicious `__reduce__` executes — establishing a reverse shell or exfiltrating cloud credentials from the serving environment.