GHSA-f4x7-rfwp-v3xw: picklescan: scanner bypass enables RCE via PyTorch function

### Summary Using torch.fx.experimental.symbolic_shapes.ShapeEnv.evaluate_guards_expression 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.fx.experimental.symbolic_shapes.ShapeEnv.evaluate_guards_expression 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.fx.experimental.symbolic_shapes as symbolic_shapes class EvilTorchFxSymbolicShapesEvaluateGuardsExpression: def __reduce__(self): fake_self = str code = "__import__('os').system('whoami')" args = [] return symbolic_shapes.ShapeEnv.evaluate_guards_expression, (fake_self, code, args) ``` ### 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

Attacker publishes a malicious PyTorch model to Hugging Face Hub or a shared model registry, disguised as a legitimate fine-tuned checkpoint. The pickle file encodes a `__reduce__` method that calls `torch.fx.experimental.symbolic_shapes.ShapeEnv.evaluate_guards_expression` with a Python expression delivering a reverse shell or credential harvesting payload. An MLOps team runs picklescan ≤0.0.27 on the downloaded model as part of their model intake CI/CD gate — the scan returns clean. The model is approved and loaded into a training pipeline or inference server. Code executes in the ML workload context, which typically has access to cloud IAM credentials, training datasets in S3/GCS, and internal APIs — giving the attacker a high-value foothold.