Korean AI startup VIDRAFT's "ViKernel" moves AI security from the perimeter into the model itself — a fundamental shift in how enterprise AI agents can be controlled.
TL;DR: VIDRAFT, a Korean AI startup, has unveiled ViKernel, a security operating system embedded directly inside AI model weights, designed to prevent unauthorized data exfiltration, mask sensitive personal information, and halt anomalous agent behavior via a kill switch. In internal testing across 661 trials using four publicly available AI models, ViKernel recorded zero unauthorized external data transmissions. VIDRAFT has launched a public demo on Hugging Face and plans to position the technology for compliance with South Korea's AI Basic Act and the EU AI Act.
Korean AI startup VIDRAFT announced on November 3, 2026, that it has developed ViKernel — a security operating system that operates from within AI model weights rather than as an external safeguard — marking a significant architectural departure from conventional enterprise AI security approaches. The company says the technology can block unauthorized data exfiltration, redact personal identifiers before they leave the model, detect abnormal agent behavior, and maintain a tamper-evident audit log, all without relying on external monitoring layers.
VIDRAFT's ViKernel is designed to embed security functionality directly into the numerical weight data that stores an AI model's learned behavior. The company explains that conventional AI security typically inspects inputs, outputs, and system-level access from outside the model. By fusing security logic with the weights themselves, VIDRAFT says the control mechanisms travel with the model even if it is migrated to a different server — a property the company describes as a core advantage over perimeter-based approaches.
ViKernel evaluates actions an AI agent intends to take against a predefined set of security rules before execution. Unauthorized outbound data transfers are blocked at the source. The system can also redact personally identifiable information — such as national ID numbers and bank account details — before that data can be transmitted externally.
If anomalous behavior is detected, a kill switch halts the AI agent's operation entirely. Restarting a suspended system requires an operator-signed authentication key, ensuring that a stopped agent cannot be trivially reactivated. All attempted, permitted, and blocked actions are logged, and each log entry is chained using cryptographic hashing so that any retroactive alteration is detectable during verification.
VIDRAFT says it validated ViKernel's data-exfiltration prevention in internal testing: across 661 trials using four publicly available AI models, simulated tools, and synthetic data, no unauthorized external transmissions occurred. The company itself acknowledges that these results reflect internal testing conditions, and that independent verification in real-world enterprise environments with varied attack vectors would be required to draw broader conclusions.
The technical implementation leverages circuits built within the AI model that perform computational and memory functions. VIDRAFT reports applying the technology to its own roughly 27-billion-parameter model as well as to models from multiple external model families, including Google Gemma and Alibaba Qwen. As one demonstration of computational precision, the company says experiments included multiplying two 300-digit numbers to produce a correct 600-digit result — illustrating the arithmetic accuracy of in-weight computation that underpins ViKernel's rule evaluation and logging capabilities.
A public demo called "ViKernel Desktop" is available on Hugging Face, presenting a Windows-like interface where users can explore the kill switch, data-exfiltration blocking, personal-information masking, and audit-log verification features. VIDRAFT notes, however, that the demo replays previously generated responses rather than running the security OS in real-time inference, meaning it is not a live end-to-end validation environment.
VIDRAFT CEO Kim Min-sik stated: "For enterprises to trust and use AI, blocking, recording, and stopping must all happen in one place — and that place has to be inside the model, so that security travels with the weights wherever they go."
Enterprise adoption of AI agents has accelerated regulatory scrutiny. The EU AI Act requires high-risk AI systems to automatically log operational records, and South Korea's AI Basic Act introduces its own compliance obligations. VIDRAFT explicitly frames ViKernel as a compliance tool for both regulatory frameworks, though the company and observers note that possessing an audit-log capability alone does not automatically satisfy all requirements under either law.
The startup is also a participant in a national cybersecurity-focused AI foundation model development consortium that includes Naver Cloud. Within that program, VIDRAFT is responsible for AI model safety diagnosis and evaluation, and the company plans to evolve ViKernel into a broader security management layer for enterprise AI agents. The move signals that in-weight security is being taken seriously at a government-program level, not merely as a commercial product pitch.
Q: What is VIDRAFT's ViKernel, and how does it differ from standard AI security?
A: ViKernel is a security operating system embedded within an AI model's weight data rather than installed as an external layer. Unlike conventional approaches that monitor inputs and outputs from outside the model, ViKernel enforces security rules, logs actions, and can halt agent execution from within the model itself, so controls remain intact even if the model is relocated to a different server.
Q: How did ViKernel perform in VIDRAFT's internal testing?
A: Across 661 trials using four publicly available AI models, simulated tools, and synthetic data, ViKernel recorded zero unauthorized external data transmissions. VIDRAFT cautions that these are internal results and that further independent verification in real enterprise environments is needed.
Q: Can ViKernel help companies meet EU AI Act compliance requirements?
A: VIDRAFT is positioning ViKernel partly as a regulatory compliance tool, citing its tamper-evident audit log as relevant to the EU AI Act's requirement that high-risk AI systems automatically record operational activity. However, VIDRAFT and independent analysts note that having an audit-log feature does not by itself satisfy all applicable regulatory requirements.
Source: ZDNet코리아 (2026-11-03) — original article