CVE-2026-116608 Details

A vulnerability in DAMO Academy VideoLLaMA2 versions through 1.1 allows for improper validation of integrity check values during model loading. The issue arises in the 'load_pretrained_model' function within 'videollama2/model/__init__.py'. When a pretraining checkpoint is loaded without an explicit 'model_base' argument, the application silently redirects model loading to an arbitrary repository based on the '_name_or_path' field in the checkpoint's 'config.json'. This behavior can be exploited to download and instantiate attacker-chosen tokenizer and model weight files, potentially leading to backdoored model outputs or manipulated benchmark results.

Users should avoid loading untrusted pretraining checkpoints without an explicit 'model_base' argument. If the vulnerability is confirmed, the package maintainers should update the model loading function to validate the integrity of downloaded files and ensure that '_name_or_path' does not dictate the origin of fetched resources.