High · published December 27, 2025
🚨 Watch out! An insecure deserialization vulnerability in LMDeploy could let attackers execute arbitrary code just by loading a malicious model file. ⚡ Think of this as a poorly secured package delivery service that opens any parcel without checking the contents first—one malicious model could unleash chaos right on your machine! 📦 If exploited, an attacker could run any code they want on your system, potentially compromising sensitive data or taking full control of your environment. This could lead to catastrophic data breaches or manipulation of your deployed models, making it a serious threat!
Think of this as a poorly secured package delivery service that opens any parcel without checking the contents first—one malicious model could unleash chaos right on your machine! 📦 This vulnerability arises because the `torch.load()` function is called without the `weights_only=True` parameter when loading model checkpoints, allowing an attacker to inject malicious code through crafted .bin or .pt files.
If exploited, an attacker could run any code they want on your system, potentially compromising sensitive data or taking full control of your environment. This could lead to catastrophic data breaches or manipulation of your deployed models, making it a serious threat! To protect yourself, upgrade LMDeploy to version 0.11.1 immediately. Additionally, ensure that you're validating all model files thoroughly before loading them, especially from untrusted sources. Don't forget to inform your team about this issue to spread awareness! You’ve got this! Stay vigilant and follow these steps to safeguard your deployment. 🛡️
| Product | Versions | Fixed in |
|---|---|---|
| internlm/lmdeploy | < 0.11.1 | 0.11.1 |