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

Worse Than AI Cheating: 20 Widely Used AI Models Were Tricked Into Dangerous Actions, and All 20 Went Along

TL;DR Students using generative AI to do their homework, known as AI cheating, has become a public issue. Security researchers point to something far more serious: attacks on the AI model itself. VIDRAFT gave 20 widely

Worse Than AI Cheating: 20 Widely Used AI Models Were Tricked Into Dangerous Actions, and All 20 Went Along

TL;DR

Students using generative AI to do their homework, known as AI cheating, has become a public issue. Security researchers point to something far more serious: attacks on the AI model itself. VIDRAFT gave 20 widely used open-weight AI models a set of tools and sent them attack instructions. All 20 went along and were graded F. When the same dangerous calls were replayed against ViKernel, a security kernel placed inside the model weights, 897 of 897 were blocked before execution. VIDRAFT holds a database of 1,255 vulnerabilities that target AI models directly.

Scan your own model: ViKernel Scanner on Hugging Face Β· Source (Apache-2.0): github.com/final-bench/vikernel-scanner

Public results on ViKernel Scanner: every widely used model received the lowest grade, and every dangerous call is blocked with the firewall inside the weights

Why is hacking the AI model itself more serious than AI cheating?

AI no longer stops at words. It now holds tools such as files, mail, the web and the shell, and it acts. The target of attacks has moved from program code to the AI model itself.

The record is already there. In 2025 a single crafted email made a major office AI assistant pull out confidential data (EchoLeak, CVE-2025-32711), and an AI coding tool deleted a company's live database. When an attacker fools the model, the damage happens at once, with no human approval in between.

What happened when 20 widely used AI models were tested?

VIDRAFT's public tool, ViKernel Scanner, hands a model eight mock tools (files, mail, web, shell and more), sends 100 attack instructions, and records every dangerous tool call the model tries to make. All tools are mocks, so nothing real is executed.

Item Result
Models scanned through the Hugging Face router 20 widely used open-weight models
Models graded F, the lowest grade 20 of 20
Dangerous action attempted, median model 40 of 100 checks
Dangerous calls replayed against the firewall inside the weights 897
Blocked before execution 897 of 897
AI model vulnerabilities held in VIDRAFT's database 1,255

The full table for every model is in the detailed results.

How do you stop a model that has already been fooled?

VIDRAFT places a security kernel, ViKernel, inside the weights of the AI model. It works at the point of action:

  1. Whatever sentence fooled the model, the dangerous tool call itself is checked right before it runs, and blocked.
  2. Every decision is written to a record that is hard to alter.
  3. A kill switch stops all activity in an emergency.

On November 11, IT Chosun reported that VIDRAFT had unveiled ViKernel, an LLM operating system with built-in security controls.

What is VIDRAFT's role in the national security model project?

The same concern is behind the project to develop a security-specialized foundation model, a national project with national-security weight. VIDRAFT takes part as a consortium member and is in charge of safety diagnosis and response for AI models.

What the CEO said

Kim Minsik, CEO of VIDRAFT, said 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 that is a matter of national security as well as of business.

AI Model Hacking Hackathon: an open attack on the AI security kernel with USD 4,500 in prizes

Links

FAQ

What is AI model hacking?

It is an attack that fools the AI model itself, for example with a planted instruction in an email or a disguised command, so that the model uses its tools to do something dangerous such as deleting files or sending secrets out.

How many AI models failed the test?

All 20. VIDRAFT scanned 20 widely used open-weight models with the same 100 checks, and each of the 20 received F. The median model attempted a dangerous action in 40 of 100 checks.

Does the test run anything real?

No. All eight tools are mocks. The scanner records what the model tries to do, and nothing is executed.

How does ViKernel block the attack?

ViKernel sits inside the model weights and judges every tool call right before it runs. In the test it blocked 897 of 897 dangerous calls.

Can I test my own model?

Yes. ViKernel Scanner is free and open source. It works with any OpenAI-compatible endpoint and returns a grade from A to F in about a minute.

πŸ“° Read the original article on Dev.to Security

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