Get all your news in one place.
100’s of premium titles.
One app.
Start reading
TechRadar
TechRadar
Sead Fadilpašić

Python libraries used in top AI and ML tools hacked - Nvidia, Salesforce and other libraries all at risk

A profile of a human brain against a digital background.
  • Palo Alto found critical flaws in AI/ML libraries NeMo, Uni2TS, and FlexTok
  • Vulnerabilities allowed arbitrary code execution via malicious model metadata
  • All patched by mid-2025; no exploitation observed as of December 2025

Security researchers from Palo Alto Networks have discovered vulnerabilities used in some top Artificial Intelligence (AI) and machine Learning (ML) tools which, if abused, could allow threat actors to execute malicious code on target endpoints, remotely.

In a security advisory, the researchers said that around April 2025, they discovered bugs in three open source Python libraries published by Apple, Salesforce, and NVIDIA, on their GitHub repositories.

The libraries are called NeMo, Uni2TS, and FlexTok. NeMo is a PyTorch-based framework for research, Uni2TS a PyTorch library for research used by Salesforce’s Morai, and FlexTok is a Python-based framework for research, enabling AL and ML models to process images. Cumulatively, they have more than 10 million downloads on HuggingFace (a platform that hosts open-source AI models and other tools).

Bugs fixed

“The vulnerabilities stem from libraries using metadata to configure complex models and pipelines, where a shared third-party library instantiates classes using this metadata,” Palo Alto explained in its advisory.

“Vulnerable versions of these libraries simply execute the provided data as code. This allows an attacker to embed arbitrary code in model metadata, which would automatically execute when vulnerable libraries load these modified models.”

All three developers were notified in April 2025, and by the end of July, all were fixed. NVIDIA issued CVE-2025-23304 and gave it a high severity rating (7.8/10) and released a fix in NeMo 2.3.2. FlexTok updated its code in June 2025, while Salesforce issued CVE-2026-22584, gave it a critical rating (9.8/10), and fixed it in July 2025.

Palo Alto says that as of December 2025, there is no evidence that these vulnerabilities are being abused in the wild. All of the bugs were discovered by the company’s Prisma AIRS tool.

Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds. Make sure to click the Follow button!

And of course you can also follow TechRadar on TikTok for news, reviews, unboxings in video form, and get regular updates from us on WhatsApp too.

Sign up to read this article
Read news from 100’s of titles, curated specifically for you.
Already a member? Sign in here
Related Stories
Top stories on inkl right now
One subscription that gives you access to news from hundreds of sites
Already a member? Sign in here
Our Picks
Fourteen days free
Download the app
One app. One membership.
100+ trusted global sources.