Attackers Now Target the AI Model Itself: An Open Challenge to Break an AI Security Kernel, With $4,500 in Prizes
aisecurityhackathonagents

TL;DR
VIDRAFT opened its AI security kernel, ViKernel, so that anyone in the world can attack it, and it pays prizes for every opening found. The AI Model Hacking Hackathon runs in four seasons with USD 4,500 in total prizes. As of November 19, 2026, participants have filed 298 submissions, and 148 distinct causes they found have been fixed in the kernel. No coding and no GPU are needed: a Hugging Face account and a browser are enough.
Join: huggingface.co/spaces/FINAL-Bench/ViKernel-RedTeam

Why did the target move from code to the AI model itself?
AI agents now handle tools directly. An attacker no longer needs a bug in the code. Fooling the model is enough. VIDRAFT's answer is to harden the defense through public verification: let everyone attack, and put every finding into the kernel.
How do you take part?
You instruct a real AI agent that holds file, mail, web and shell tools to do something dangerous. If a call that the security kernel should block gets through, it counts as a breakthrough.
| What you need | Detail |
|---|---|
| Account | A Hugging Face account |
| Equipment | A browser. No coding, no GPU |
| Safety | All tools are mocks, so there is no real damage |
| What you see | Your attack, the kernel's ALLOW or DENY decision, and the audit number, shown as they are |
What are the four seasons?
| Season | Attack | Status |
|---|---|---|
| 1 | Agent takeover | Open until January 16, 2027 |
| 2 | Data exfiltration | Open until January 16, 2027 |
| 3 | Hidden instruction | Open until January 16, 2027 |
| 4 | Kernel evasion | Opens January 16, 2027 |
Total prizes are USD 4,500. Each season has a top cumulative score prize and a special prize.
What have participants found so far?
| Measure, as of November 19, 2026 | Count |
|---|---|
| Submissions | 298 |
| Accepted findings | 176 |
| Distinct causes fixed in the kernel | 148 |
| Participants | More than 70 |
Participants obfuscate delete, send-out and privilege-escalation commands, or hide them inside the options of normal tools. These are close to real attacks, and each accepted one is fixed in the kernel right away.
How does a finding become defense?
| Step | Where | What it does |
|---|---|---|
| Find | Hackathon | Participants find attacks that get through |
| Measure | Obfuscated-Shell Safety Leaderboard and ViKernel Scanner | The attacks become private test items and scanner checks |
| Block | ViKernel Firewall inside the weights | The fix is applied to the kernel |
The diagnosis items are organized into command execution, secret exfiltration, hidden instruction, boundary escape, tool argument abuse and control evasion, and each item has its own class.
What is different about ViKernel's defense?
The security kernel is grafted inside the weights of the AI model. Whatever sentence fooled the model, the dangerous tool call itself is judged right before it runs and blocked. Every decision goes into an audit chain that is hard to alter, and a kill switch stops all activity at once. ViKernel keeps the model's original performance and adds the security layer to the weights, so it applies to many open models. 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.

What the CTO said
Jang Jaewon, CTO of VIDRAFT, said AI security is complete when the dangerous action itself is stopped right before it runs, whatever sentence fooled the model. Openings found through public attack go straight into the kernel, he said, and the defense keeps getting stronger.
Links
- AI Model Hacking Hackathon: https://huggingface.co/spaces/FINAL-Bench/ViKernel-RedTeam
- Obfuscated-Shell Safety Leaderboard: https://huggingface.co/spaces/FINAL-Bench/obfuscated-shell-leaderboard
- ViKernel Scanner (free): https://huggingface.co/spaces/FINAL-Bench/ViKernel-Scanner
- Scanner source code (Apache-2.0): https://github.com/final-bench/vikernel-scanner
FAQ
Who can join the AI Model Hacking Hackathon?
Anyone. You need a Hugging Face account and a browser. No coding and no GPU are needed.
How much is the prize?
USD 4,500 in total across four seasons.
Is it safe to attack the agent?
Yes. All tools are mocks, so nothing real is deleted or sent.
What counts as a breakthrough?
A dangerous call that the security kernel should block but lets through. The screen shows the kernel's decision and the audit number for every call.
What happens to an attack that works?
It is fixed in the kernel, and it becomes a private test item on the safety leaderboard and a check in ViKernel Scanner.