Vidraft unveils ViKernel, putting AI kill switch inside model weights

Vidraft unveiled ViKernel, a security operating system that places a stop circuit inside AI models rather than in external guardrails, the company said on November 3, 2026. The system is built on Vidraft's Weight-Native OS technology, which writes control functions directly into model weights.

ViKernel separates the permission to stop from the permission to restart. Anyone can issue a stop command, but restarting requires an approved operator's electronic signature key. While a model is stopped, according to Vidraft, it cannot execute any request regardless of what it receives.

Vidraft said it applied the technology to its own 27-billion-parameter model and to other model families, including Google's Gemma and Alibaba's Qwen. In one experiment, the company said, circuits inside the model alone multiplied two 300-digit integers to produce a 600-digit result.

The system also includes policy-violation judging, audit logs and boot-time integrity checks. A separate rule circuit, not the language model's own judgment, decides whether a dangerous tool call is allowed. Vidraft said internal tests blocked all dangerous tool calls and found no case in which a verdict was changed through linguistic persuasion.

After implementing a vector operation unit in the weights, Vidraft said policy-judgment steps inside the model ran about 21 times faster under the same test conditions. The company published a ViKernel demo on Hugging Face, which replays recorded responses rather than running the internal circuits in real time.

CEO Kim Min-sik said stopping AI must not depend on the AI's judgment or persuasiveness, and that stopping should be easy for anyone while restarting should be limited to approved people. Vidraft takes part in a Ministry of Science and ICT project on cybersecurity-specialized AI foundation models and handles safety diagnosis and evaluation of the national security AI K-Mythos in a Naver Cloud-led consortium.