Executive brief
CosyVoice is a voice generation model used for creating and processing speech data. A security flaw in its data processing tool allows an attacker to execute malicious code on a user's computer if the user is tricked into processing a specially crafted data file. This could lead to a full system compromise, unauthorized data access, or the installation of malware.
Technical details
An insecure deserialization vulnerability (CWE-502) exists in the 'make_parquet_list.py' tool within CosyVoice. The vulnerability stems from the use of 'torch.load()' to process PyTorch (.pt) files—such as utterance and speaker embeddings—without the 'weights_only=True' parameter enabled. Because 'torch.load' uses the Python pickle module by default, it can be coerced into deserializing arbitrary Python objects. An attacker can exploit this by providing a malicious .pt file in a data directory; when a victim runs the processing script on that directory, the attacker's code is executed with the privileges of the user running the script.
Affected products
- FunAudioLLM CosyVoice up to commit 6e01309e01bc93bbeb83bdd996b1182a81aaf11e
Timeline
- 2026-05-11: advisory: CVE-2026-31249 published by NVD