Lightx2V
Monthly
Code injection in ModelTC LightX2V through version 0.5.0 allows a local attacker to execute arbitrary code when a victim loads an attacker-controlled tokenizer/model via the cosmos3 runner's AutoTokenizer.from_pretrained call, by supplying a crafted auto_map value in the model configuration. Exploitation requires the user to load untrusted model content and runs only with the local user's existing privileges; there is no remote network attack surface. Publicly available exploit code exists, the project has not responded or released a patch, and the assessed real-world risk remains low because the attack is local and depends on the victim choosing to load a malicious model.
Code injection in ModelTC LightX2V through version 0.5.0 allows a local attacker to execute arbitrary code when a victim loads an attacker-controlled tokenizer/model via the cosmos3 runner's AutoTokenizer.from_pretrained call, by supplying a crafted auto_map value in the model configuration. Exploitation requires the user to load untrusted model content and runs only with the local user's existing privileges; there is no remote network attack surface. Publicly available exploit code exists, the project has not responded or released a patch, and the assessed real-world risk remains low because the attack is local and depends on the victim choosing to load a malicious model.