A newly disclosed critical vulnerability (CVE-2025-32434) in PyTorch, the widely used open-source machine learning framework, allows attackers to execute arbitrary code on systems loading AI models—even when safety measures like weights_only=True are enabled.
The flaw impacts all PyTorch versions ≤2.5.1 and has been patched in version 2.6.0, released earlier this week.
CVE ID | Severity | Affected Versions | Patched Version |
CVE-2025-32434 | Critical | PyTorch ≤2.5.1 (pip) | 2.6.0 |
The flaw resides in PyTorch’s torch.load() function, which is commonly used to load serialized AI models.
While enabling weights_only=True was previously believed to prevent unsafe code execution by restricting data loading to model weights, security researcher Ji’an Zhou demonstrated that attackers can bypass this safeguard to execute remote commands.
This undermines a core security assumption in PyTorch’s documentation, which explicitly recommended weights_only=True as a mitigation against malicious models.
Organizations using this setting to protect inference pipelines, federated learning systems, or model hubs are now at risk of remote takeover.
The PyTorch team acknowledged the vulnerability, stating, “This issue highlights the evolving nature of ML security. We urge all users to update immediately and report suspicious model behavior.”
PyTorch is foundational to AI research and deployment, with users ranging from startups to tech giants like Meta and Microsoft.
This vulnerability exposes critical infrastructure to attacks that could steal data, disrupt services, or hijack resources.
As AI adoption grows, securing model pipelines is paramount. CVE-2025-32434 serves as a stark reminder that even trusted safeguards require continuous scrutiny.
Update PyTorch installations, audit model sources, and treat all third-party AI artifacts as potential attack vectors until verified.
Find this News Interesting! Follow us on Google News, LinkedIn, & X to Get Instant Updates!
A new project has exposed a critical attack vector that exploits protocol vulnerabilities to disrupt…
A threat actor known as #LongNight has reportedly put up for sale remote code execution…
Ivanti disclosed two critical vulnerabilities, identified as CVE-2025-4427 and CVE-2025-4428, affecting Ivanti Endpoint Manager Mobile…
Hackers are increasingly targeting macOS users with malicious clones of Ledger Live, the popular application…
The European Union has escalated its response to Russia’s ongoing campaign of hybrid threats, announcing…
Venice.ai has rapidly emerged as a disruptive force in the AI landscape, positioning itself as…