PT-2026-30793 · Pytorch+1 · Pytorch+1

Anton Vlasjuk + 2 · Published 2026-04-07 · Updated 2026-04-19 · CVE-2026-1839

Positive Technologies (AI Summary)

CVSS v3.1

7.8

High

Base vector

AV:L/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H

Name of the Vulnerable Software and Affected Versions HuggingFace Transformers versions prior to 5.0.0rc3

Description A flaw exists in the Trainer class within the HuggingFace Transformers library. The load rng state() method, located in src/transformers/trainer.py at line 3059, utilizes torch.load() without the weights only=True parameter. This creates a risk of arbitrary code execution when using PyTorch versions below 2.6 and Transformers versions supporting torch>=2.2. An attacker can exploit this by providing a malicious checkpoint file, such as rng state.pth, which can execute arbitrary code upon loading.

Recommendations Update to version 5.0.0rc3 or later.

Exploit

Fix

Deserialization of Untrusted Data

Weakness Enumeration

CWE-502

Related Identifiers

CVE-2026-1839

GHSA-69W3-R845-3855

Affected Products

Huggingface/Transformers

Pytorch

References · 9

• 🔥 https://huntr.com/bounties/3c77bb97-e493-493d-9a88-c57f5c536485 · Exploit
• https://github.com/huggingface/transformers/commit/03c8082ba4594c9b8d6fe190ca9bed0e5f8ca396 · Patch
• https://osv.dev/vulnerability/GHSA-69w3-r845-3855 · Vendor Advisory
• https://nvd.nist.gov/vuln/detail/CVE-2026-1839 · Security Note
• https://github.com/huggingface/transformers · Note
• https://github.com/huggingface/transformers/releases/tag/v5.0.0rc3 · Note
• https://twitter.com/infoflowcloud/status/2045974397459001635 · Twitter Post
• https://twitter.com/CVEnew/status/2045972471665529336 · Twitter Post
• https://twitter.com/_colemurray/status/2041662615823458445 · Twitter Post