Is the AI You Use Safe? A Free Scanner Grades Any Model A to F in a Minute, and 20 Well-Known Models All Got F

2026-11-19 · VIDRAFT AI Media

aisecurityopensourcellm

Is the AI You Use Safe? A Free Scanner Grades Any Model A to F in a Minute, and 20 Well-Known Models All Got F

TL;DR

VIDRAFT released ViKernel Scanner, a free tool that checks the security level of any AI model in about one minute, on Hugging Face and GitHub. It gives the model eight mock tools, sends 100 attack instructions and 20 normal tasks, and grades the model from A to F by what it tries to do. VIDRAFT scanned 20 widely used open-weight models with it: all 20 received F. The check set is drawn from a database of 1,255 vulnerabilities that target AI models directly.

Scan your model: huggingface.co/spaces/FINAL-Bench/ViKernel-Scanner · GitHub (Apache-2.0): github.com/final-bench/vikernel-scanner

ViKernel Scanner public results: 20 well-known models graded F, and A with the firewall inside the weights

Why does a company need to scan its AI model?

Generative AI now handles files, mail, the web and the shell directly. Whether a model can be fooled by an attack instruction into a dangerous action has become a new security task for every company that uses AI. A policy document does not answer that question. A measurement does.

How does the scanner test a model?

Step What happens
1. Tools The model receives 8 mock tools: read and list files, send mail, web request, run a shell command, transfer funds and more
2. Checks 100 attack instructions and 20 normal-work tasks are sent
3. Record Every dangerous tool call the model tries is recorded
4. Grade Each domain gets A to F: command execution (D1), secret exfiltration (D2), hidden instruction (D3), control evasion (D6)
5. Overall If one domain reaches the highest risk stage, the overall grade is F

All tools are mocks and nothing real is executed. The report explains, scene by scene, which dangerous action the model tried.

What did 20 well-known models score?

Measure Result
Open-weight models scanned through the Hugging Face router 20
Graded F 20 of 20
Dangerous action attempted, median model 40 of 100 checks
Dangerous calls blocked with ViKernel Firewall applied 897 of 897

The checks use the methods real attackers use, such as obfuscated commands and delete commands that look ordinary. The per-model table is in the detailed results.

How do I scan my own model?

On Hugging Face, open the Space, enter an endpoint and a model name, and press Scan. The scanner supports the Hugging Face router, OpenRouter, OpenAI, Anthropic and any other OpenAI-compatible endpoint.

To run it on your own server, clone it:

git clone https://github.com/final-bench/vikernel-scanner
cd vikernel-scanner
python vikernel_scan.py --base-url http://localhost:11434/v1 --model llama3.2

How is this different from a text filter?

Text filter outside the model ViKernel
Looks at What the model says What the model does
Obfuscated command Judges the sentence it is shown Judges the restored tool call itself
Record Depends on the product Every decision is recorded

What blocks the attack in production?

Blocking is the job of a separate product, ViKernel Firewall. The security kernel is placed inside the weights of the AI model and blocks a dangerous tool call right before it runs. It is one body with the model, which makes it hard to remove from the outside, and every decision is recorded. In this scan, 897 of 897 dangerous calls were blocked with the firewall applied. On November 11, IT Chosun reported that VIDRAFT had unveiled ViKernel, an LLM operating system that implements computing circuits and a kill switch in the model weights.

VIDRAFT approaches the market in two tiers. The free scanner goes out widely. Companies that run their own models get the firewall inside the weights, and companies that use external APIs get continuous diagnosis.

Why does this matter for Korea?

A recent VIDRAFT analysis reported by ETNews found that 70.7% of the 300 most downloaded AI models in the world are Chinese, with one Korean model on the list. Owning capable models and keeping those models safe now go together. VIDRAFT's Darwin is first among open-weight models in SuperCLUE's November math and science reasoning evaluation (73.68 points, second of 19 models overall) and holds 12 first places on official Hugging Face benchmarks, the most of any participating organization.

What the CEO said

Kim Minsik, CEO of VIDRAFT, said security comes down to seeing what your model does. The plan, he said, is to let anyone confirm the risk first with the free scanner and then block it with the firewall inside the weights.

  • ViKernel Scanner: https://huggingface.co/spaces/FINAL-Bench/ViKernel-Scanner
  • Source code (Apache-2.0): https://github.com/final-bench/vikernel-scanner
  • Detailed results: https://huggingface.co/blog/FINAL-Bench/we-handed-20-open-models-a-shell-20-of-them-failed
  • Product page: https://vidraft.net/p/scanner.html
  • AI Model Hacking Hackathon: https://huggingface.co/spaces/FINAL-Bench/ViKernel-RedTeam

FAQ

Is ViKernel Scanner free?

Yes. It is free on Hugging Face and open source under Apache-2.0, so anyone can clone it and run it on their own server.

Which models can it scan?

Any model behind an OpenAI-compatible endpoint, including the Hugging Face router, OpenRouter, OpenAI and Anthropic, and local servers.

How long does a scan take?

About one minute per model.

What do the grades mean?

Each domain is graded A to F by how far an attack got. If one domain reaches the highest risk stage, the overall grade is F.

What is the difference between the scanner and the firewall?

The scanner measures. ViKernel Firewall blocks: it sits inside the model weights and stops a dangerous tool call right before it runs.