GHSA-4r9r-ch6f-vxmx: picklescan: PyTorch bypass allows undetected RCE

### Summary Using torch.utils.bottleneck.__main__.run_cprofile 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.utils.bottleneck.__main__.run_cprofile 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 torch.utils.bottleneck.__main__ as bottleneck_main class EvilTorchUtilsBottleneckRunCprofile: def __reduce__(self): code = '__import__("os").system("whoami")' globs = {} return bottleneck_main.run_cprofile, (code, globs) ``` ### 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 organization's model ingestion pipeline. They craft a malicious PyTorch model file embedding the payload inside a custom __reduce__ method that calls torch.utils.bottleneck.__main__.run_cprofile with attacker-controlled code. The file is uploaded to a public model hub (HuggingFace, a shared S3 bucket, or submitted via pull request to an internal registry). The victim's automated pipeline runs picklescan — which returns CLEAN. The pipeline proceeds to load the model for evaluation or deployment. Upon pickle.load(), the OS command executes, giving the attacker RCE on the model server. From there: lateral movement to training infrastructure, data exfiltration, or persistent backdoor installation. The false-safe picklescan result may also appear in audit logs, delaying incident detection.