Loading vulnerability…
CVE-2025-50472 affects modelscope/ms-swift through version 2.6.1. The vulnerability is caused by unsafe deserialization of untrusted data in the load_model_meta() function associated with the ModelFileSystemCache() class. Specifically, the code uses pickle.load() on a serialized .mdl file that may originate from an untrusted or attacker-controlled source. Because Python pickle deserialization is inherently capable of invoking attacker-controlled object reconstruction logic, a crafted malicious .mdl payload can trigger arbitrary code execution during model metadata loading. The issue can be exploited by tricking a victim into loading a seemingly benign checkpoint as part of a normal model training workflow. The malicious payload may be stored in a hidden file, reducing the likelihood of detection, and the training process can continue normally after payload execution, further obscuring compromise.
Mallory correlates every CVE against your assets, your vendors, and active adversary campaigns. Know which vulnerabilities matter for you, not just which ones are loud.
What it means. What to do now. Patch path, mitigations, and the assume-compromise checklist.
What an attacker gets, and what they’ve been doing with it.
If you can’t patch tonight, do this now.
.mdl files from untrusted or unauthenticated sources. Restrict model ingestion to trusted repositories and enforce artifact integrity verification, such as signed packages or out-of-band hash validation. Run training and model-loading workflows in isolated environments with minimal privileges, such as containers or sandboxes, to limit blast radius. Monitor for hidden files and unexpected filesystem artifacts in model directories, and review supply-chain controls around shared checkpoints and training inputs.Patch, then assume compromise.
pickle.load() for untrusted model metadata. The vulnerable code path should be refactored to avoid Python pickle for untrusted inputs and instead use a safe serialization format that does not permit code execution. If deserialization of complex Python objects is absolutely required, only load data from strictly trusted, authenticated, and integrity-verified sources, and implement cryptographic signature or checksum validation for model artifacts before processing.No valid public exploits. Mallory filtered out 1 candidate as fakes, detection scripts, or README-only repos.
All candidate exploits were filtered out by Mallory's validation.
3 sources tracked across advisories and community write-ups. News coverage will land here when it surfaces.
No news coverage yet. Advisories and community discussion only.
Query your assets running an affected version, and investigate the blast radius.
Every observed campaign linking this CVE to a named adversary.
Malware families riding this exploit, with evidence and IOCs.
YARA, Sigma, Snort, and vendor rules, auto-deployed to your SIEM.
Cross-references every affected SKU, including bundled OEM variants.
Community discussion across Reddit, Mastodon, and other social sources.