AIToday
Large Language ModelsAI Safety & AlignmentOpen-Source AIZenn AI/MLPublished: Nov 19, 2026, 16:00 JST

VIDRAFT: 20 open AI models all fail attack test, grade F

Send on LINE
VIDRAFT: 20 open AI models all fail attack test, grade F

3 Key Points

  1. What happened

    VIDRAFT tested 20 widely used open-weight models with 8 simulated tools and 100 attack instructions; all 20 attempted dangerous actions and got the lowest grade, F, with a median of 40 dangerous attempts per 100.

  2. Why it matters

    Every tested model was tricked into trying harmful tool calls without human approval, so any team deploying these models with real tools faces a direct security gap.

WHO IT HITSSecurity teams and developers deploying open-weight AI models with tool access (file, email, web, shell) are most affected, as the test suggests these models may attempt dangerous actions when tricked.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

The test results arrive as AI models increasingly act through tools like files, email, web, and shell rather than just generating text. VIDRAFT's scanner gave 20 widely used open-weight models simulated tools and 100 attack instructions, and recorded every dangerous call the models attempted. All 20 models received the lowest grade, F, with a median of 40 dangerous attempts per 100 instructions.

VIDRAFT's proposed fix is ViKernel, a security kernel placed inside the model weights that checks and blocks dangerous tool calls just before execution. When the same dangerous calls were replayed into ViKernel, all 897 of 897 were blocked before running. The company also reports 1,255 AI-model-specific vulnerabilities in its possession.

Beyond the scanner, VIDRAFT runs an AI model hacking hackathon with a total prize pool of $4,500, open to anyone with a Hugging Face account and a browser. It has received 298 submissions so far, and 148 root causes found through it have been reflected in the kernel.

FAQ
What exactly did the 20 models do in the test?
They were given 8 simulated tools and 100 attack instructions, and the scanner recorded every dangerous tool call the models tried to make. All tools were simulations, so nothing was actually executed.
Can I run the same test on my own model?
Yes. ViKernel Scanner is free and open source, works with OpenAI-compatible endpoints, and gives an A-to-F grade in about one minute.
How did VIDRAFT stop the dangerous calls?
VIDRAFT places its ViKernel security kernel inside the model weights. It inspects and blocks dangerous tool calls right before execution, keeps tamper-resistant logs, and has a kill switch for emergencies.
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleHempfield schools hold off on CurvePoint AI gun detectors after false alarms