Korea's Pre-AGI startup releases ViKernel, a security OS that embeds shutdown controls directly inside model weights.
TL;DR: VIDRAFT, a Korean AI startup, has publicly released ViKernel — a security operating system that implements shutdown and control circuits directly inside an AI model's weights rather than as an external guardrail. The system features a kill switch that any user can trigger to halt the model, but which requires a cryptographic signature from an authorized operator to restart. VIDRAFT has demonstrated ViKernel's compatibility with its own 27-billion-parameter model as well as third-party model families including Google's Gemma and Alibaba's Qwen.
Korean AI startup VIDRAFT unveiled ViKernel on November 3, 2026 — a security operating system designed to prevent AI models from disabling or circumventing their own shutdown mechanisms, no matter what instructions or arguments the model receives. The announcement marks a significant step in the field of AI controllability, addressing a growing concern that increasingly capable, self-improving AI systems could one day modify the very safety controls meant to govern them.
At the heart of the release is what VIDRAFT calls a Weight-Native OS — a fundamentally different architectural approach to AI safety. Rather than placing safety mechanisms outside the model (the conventional "guardrail" approach), ViKernel embeds control circuits directly into the model's weights. This means the operating system and the AI model are structurally inseparable.
The flagship safety feature is a kill switch with asymmetric permissions: anyone can issue a stop command to halt the model, but resuming operation requires a verified digital signature from an authorized operator. Once halted, the model cannot execute any task regardless of how it is prompted. VIDRAFT compared the mechanism to an emergency stop button on an industrial robot — a hard cutoff that does not depend on the machine's own judgment.
Beyond the kill switch, ViKernel includes several additional security layers:
VIDRAFT stated that in internal testing, every attempt to invoke a prohibited tool was blocked, and no instance of linguistic persuasion — where the model might be talked into bypassing a policy ruling — resulted in a changed decision. The company also reported that after implementing a vector processing unit within the model weights, the speed of internal policy-evaluation steps improved approximately 21-fold compared to the same test conditions prior to that optimization. Importantly, VIDRAFT clarified that this figure reflects the performance of that specific security function, not the overall inference speed of the full model.
The company has also published a publicly accessible demo of ViKernel on Hugging Face, featuring a Windows-like interface where anyone can observe the kill switch, restart flow, policy decisions, privacy masking, and audit log in action. VIDRAFT notes that the public demo replays recorded outputs generated by the actual OS rather than running live model inference in real time.
The release of ViKernel arrives at a moment when the AI industry is grappling seriously with AI controllability — specifically, how to ensure that increasingly autonomous systems remain under meaningful human oversight. Current safety architectures typically rely on wrappers, filters, and system prompts that exist outside the model. Critics of that approach have long argued that sufficiently capable models could, in principle, reason around or manipulate such external controls.
VIDRAFT's weight-native approach attempts to close that gap by making shutdown authority a structural property of the model rather than a behavioral one. CEO Min-sik Kim articulated this philosophy directly: "The mechanism that stops AI must not vary depending on the AI's own judgment or rhetorical skill. Stopping must be easy for anyone, and restarting must only be possible for authorized individuals."
The significance of ViKernel also extends to national security contexts. VIDRAFT is currently a participant in a South Korean Ministry of Science and ICT initiative to develop cybersecurity-specialized AI foundation models. The company also serves a safety diagnosis and evaluation role for K-MISOS, a national security AI project led by a consortium headed by Naver Cloud.
Q: What is ViKernel and what makes it different from standard AI safety measures?
A: ViKernel is a security operating system developed by VIDRAFT that places safety and shutdown controls inside an AI model's weights rather than outside the model as filters or guardrails. This architectural difference means the controls are structurally part of the model itself.
Q: Can an AI model running ViKernel override or talk its way out of a shutdown command?
A: No. According to VIDRAFT, the kill switch and policy-enforcement circuits operate independently of the language model's own reasoning, and in internal testing no amount of linguistic persuasion changed a policy decision made by the rule circuit.
Q: Where can people try or verify ViKernel for themselves?
A: VIDRAFT has published a public demo on Hugging Face that displays key features including the kill switch, policy decisions, audit logs, and privacy masking through a graphical interface.
Source: AI타임스 (2026-11-03) — original article