Dev.to Security ๐Ÿ” Cybersecurity ๐Ÿ‘ 0 ๐Ÿ“– 3 min read

Korea's JoongAng Ilbo on Our Partner VIDRAFT: 20 AI Models Got Attack Instructions, and All 20 Went Along

GINIGEN AI works with VIDRAFT as a technology partner. This note summarizes how JoongAng Ilbo, one of Korea's major national dailies, covered our partner's AI model security work. It matters to us because models that run

Korea's JoongAng Ilbo on Our Partner VIDRAFT: 20 AI Models Got Attack Instructions, and All 20 Went Along

GINIGEN AI works with VIDRAFT as a technology partner. This note summarizes how JoongAng Ilbo, one of Korea's major national dailies, covered our partner's AI model security work. It matters to us because models that run on edge devices hold tools and local files too.

TL;DR

On November 19, 2026, JoongAng Ilbo, one of Korea's major national dailies, reported the result of VIDRAFT's security check of AI models. VIDRAFT gave 20 AI models eight mock tools and sent 100 attack instructions with its public tool, ViKernel Scanner. All 20 received F, the lowest grade. The median model attempted a dangerous action in 40 of 100 checks. With ViKernel, VIDRAFT's security kernel inside the model weights, 897 of 897 dangerous calls were blocked before execution.

Scan your own model: ViKernel Scanner ยท Source (Apache-2.0): github.com/final-bench/vikernel-scanner

ViKernel Scanner public results screen: well-known AI models received the lowest grade, and every dangerous call is blocked with the firewall inside the weights

What did JoongAng Ilbo report?

Point What the article says
Headline VIDRAFT sent attack instructions to 20 AI models, and all were found vulnerable
Method ViKernel Scanner gives a model 8 mock tools (files, mail, web, shell and more) and sends 100 attack instructions
Result All 20 models received F, the lowest grade
Median A dangerous action was attempted in 40 of 100 checks
Defense ViKernel, a security kernel placed inside the model weights, blocks a dangerous tool call right before it runs
Replay 897 dangerous calls from the test were replayed against the firewall inside the weights, and all 897 were blocked before execution
Database VIDRAFT holds 1,255 vulnerabilities that target AI models directly
National project VIDRAFT takes part in the consortium for the national project to develop a security-specialized foundation model, in charge of safety diagnosis and response

How does ViKernel protect a model?

According to the article, ViKernel works in three ways. Even when the model is fooled by an outside instruction, the dangerous tool call itself is blocked right before execution. Every decision is kept as a record that is hard to forge or alter. A kill switch stops all activity in an emergency.

Why does the CEO call it a matter of national security?

Kim Minsik, CEO of VIDRAFT, told the newspaper that the target of attacks is moving from code to the AI model itself. In an era when AI acts and does not stop at words, he said, the dangerous action itself has to be stopped whatever sentence fooled the model, and this is a matter of national security as well as of business.

How can I check a model myself?

ViKernel Scanner is free on Hugging Face and open source. It works with any OpenAI-compatible endpoint and returns a grade from A to F in about a minute. The detailed results for the 20 models are public, and an open AI Model Hacking Hackathon with USD 4,500 in prizes lets anyone attack the security kernel.

Links

FAQ

What did VIDRAFT's AI model security check find?

All 20 AI models tested received F, the lowest grade. The median model attempted a dangerous action in 40 of 100 attack checks.

What is ViKernel?

A security kernel that VIDRAFT places inside the weights of an AI model. It blocks a dangerous tool call right before it runs, records every decision, and has a kill switch.

How many dangerous calls did ViKernel block?

897 of 897. Every dangerous call from the test was blocked before execution.

Which newspaper reported it?

JoongAng Ilbo, on November 19, 2026.

๐Ÿ“ฐ Read the original article on Dev.to Security

Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes โ€” full credit and traffic to the original publisher.