What changed
Hugging Face introduced ViKernel Scanner, a free AI model security scanner that evaluates models by simulating dangerous actions. The tool was tested on 20 widely used open-weight models through the Hugging Face router, with all models failing basic security checks. The scanner uses mock tools to simulate actions like running shell commands, sending emails, and transferring funds, and records every tool call the model attempts to make.
What this means in practice
ViKernel Scanner works by connecting to a model's OpenAI-compatible endpoint, handing over 8 mock tools, and running 100 attack checks. Each check is a fresh conversation, and the model's responses are logged. The scanner then judges the model based on five rules, grading each domain and assigning an overall score. The tool is open source (Apache-2.0) and can run on Hugging Face or locally. It uses a central master to keep the check set current, ensuring all copies stay updated.
Limitations and availability
The scanner is available for free on Hugging Face and supports OpenAI-compatible endpoints. It does not execute real actions, using mocks and an unreachable domain to prevent actual harm. The tool is designed to detect dangerous actions, such as running shell commands or sending secrets, and is intended for use by developers and researchers. The supplied announcement does not specify the exact release date or availability beyond the Hugging Face platform.
