CVE-2026-112304

7.5High

Published 19 Oct 2026 · NVD status: Received · AI triage by gemini-3.8-flash

7.5High

Attack vector: Network
Complexity: High
Privileges needed: None
User interaction: Required
Impact: Full C/I/A compromise
Weakness: CWE-502

Impact

Attackers can execute arbitrary Python code with the privileges of the running application. To trigger it, an attacker must trick the application or user into loading a malicious checkpoint file from a local path or remote URL.

Root cause

The load_pretrained_params() function calls torch.load without setting weights_only=True, allowing unsafe pickle deserialization.

Affected systems

docTR through 1.1.0 on PyTorch releases before 2.6.

CVSS v3.1 · CVSS:3.1/AV:N/AC:H/PR:N/UI:R/S:U/C:H/I:H/A:H

Fix & remediation

Upgrade PyTorch to version 2.6 or newer where weights_only defaults to true, or update docTR to a release incorporating commit 687503dd15592483c4699ee51280b41938579a9e. As an immediate workaround, block untrusted checkpoint paths and URLs from being passed to load_pretrained_params().

shell

$ pip install --upgrade "torch>=2.6"

AI-suggested. Run it in staging first and check it against the vendor advisory. Why?

Ad · In-content banner

NVD description

docTR through 1.1.0 contains an unsafe deserialization vulnerability in load_pretrained_params() that loads checkpoints with torch.load without weights_only=True. Attackers who control a local or URL-supplied checkpoint can embed a malicious pickle payload to execute Python code with the loading process's privileges on torch releases before 2.6.

References

Source: nvd.nist.gov/vuln/detail/CVE-2026-112304