CVE-2026-112304: Deserialization of Untrusted Data in Mindee docTR

The vulnerability in Mindee docTR (<=1.1.0) arises from unsafe deserialization in the load_pretrained_params() function, which loads model checkpoints using torch.load without restricting to weights_only=True. This allows an attacker controlling the checkpoint source (local or URL) to craft a malicious pickle payload that executes arbitrary Python code during the loading process. The issue affects torch releases prior to version 2.6, where this unsafe loading behavior exists. The CVSS 4.0 vector indicates network attack vector, low attack complexity, partial privileges required, user interaction needed, and high impact on confidentiality, integrity, and availability.

Successful exploitation allows remote or local attackers who can supply a malicious checkpoint file to execute arbitrary Python code with the privileges of the process loading the checkpoint. This can lead to full compromise of the application environment running docTR. The vulnerability impacts confidentiality, integrity, and availability of the affected system.

No explicit patch or fix information is provided in the available data. Patch status is not yet confirmed — check the vendor advisory for current remediation guidance. Until a fix is available, avoid loading checkpoints from untrusted or unauthenticated sources. Consider upgrading torch to version 2.6 or later, where the unsafe deserialization behavior is addressed.