CVE-2026-115802 Details

A vulnerability in InternScience InternAgent version 1.5.0 allows for untrusted LoRA adapter base model names to be blindly accepted and fetched from the Hugging Face Hub. This issue is present in the 'train' function of 'tasks/AutoChem/code/experiment.py', within the LoRA Loading component. The vulnerability can be exploited remotely, leading to a form of model-provenance poisoning where the integrity of trained artifacts is compromised. While the exploitation does not result in code execution, it allows for the manipulation of model weights in a way that could subvert trust in the model's provenance.

To address this vulnerability, the 'base_model_name_or_path' field should be treated as untrusted. The application should prefer the model path provided by the operator, and implement validation to ensure consistency between the two. Any discrepancies should trigger a loud failure or warning. Additionally, the accepted values for the base model name should be restricted to a predefined allowlist of trusted sources.