HuggingFace Transformers Trainer Arbitrary Code Execution via Unsafe torch.load (CVE-2026-1839)

Vulnerability Key Information Summary 1. Vulnerability Overview Vulnerability Name: Arbitrary Code Execution via Unsafe in Trainer Checkpoint Loading in huggingface/transformers (Arbitrary Code Execution in HuggingFace Transformers Trainer Checkpoint Loading due to Unsafe ) CVE ID: CVE-2026-1839 Vulnerability Type: CWE-502: Deserialization of Untrusted Data (Deserialization of Untrusted Data) Detailed Description: In the class within , the method invokes without the parameter. Although wrapped with the context manager, this mechanism offers no protection for PyTorch versions < 2.6. Attackers can execute arbitrary code by providing a malicious checkpoint file (e.g., ) when the victim resumes training. 2. Scope of Impact Affected Component: (Line 3059) Affected Versions: All versions supporting and using . Attack Vector: Malicious checkpoint files. Severity: Medium (6.5) / High (based on Impact Assessment) Impact Assessment: Confidentiality (High): Attackers can read any file accessible by the process. Integrity (High): Attackers can modify files, install backdoors, or alter training results. Availability (High): Attackers can crash the process, delete files, or consume resources. Affected Users: HuggingFace Transformers users running PyTorch 2.2, 2.3, 2.4, or 2.5; those resuming training from untrusted checkpoints; those loading checkpoints from shared or public sources. 3. Remediation It is recommended to add the parameter to the vulnerable call. 4. POC/Exploit Code**

Goal Reached Thanks to every supporter — we hit 100%!

HuggingFace Transformers Trainer Arbitrary Code Execution via Unsafe torch.load (CVE-2026-1839)