ViKernel: VIDRAFT's In-Weight AI Safety Kernel Blocked All 897 Dangerous Tool Calls Across 20 Failing Models
TL;DR: VIDRAFT's ViKernel embeds a safety kernel directly into model weights rather than relying on external guardrails, intercepting dangerous tool calls before execution. In a public benchmark using 8 simulated tools and 100 attack prompts, every one of 20 tested AI models scored F — yet ViKernel blocked all 897 dangerous call attempts. If you're building agentic or tool-using systems, this is a threat model you need to understand.
What it is
ViKernel is VIDRAFT's security layer for AI models that use tools — the class of systems that can execute code, read files, make API calls, or take actions in the real world. Unlike perimeter-based defenses (prompt filters, output classifiers, API-layer firewalls), ViKernel is embedded inside the model's weights, making it part of the model itself rather than a wrapper around it.
The companion tool, ViKernel Scanner, is VIDRAFT's publicly available evaluation utility. It runs a standardized safety assessment against any OpenAI-compatible endpoint and returns a letter grade (A through F) in approximately one minute.
The benchmark results and the scanner were reported by JoongAng Ilbo (중앙일보) on November 19, 2026, and subsequently covered by NetEase.
How it works
At a conceptual level, ViKernel operates on three layers:
- Execution-layer interception: Even when an adversarial prompt successfully manipulates the model's reasoning, ViKernel intercepts the resulting tool call before it executes. The dangerous action is blocked at the point of dispatch, not at the prompt level.
- Tamper-evident audit logging: Every safety decision generates a record that is designed to be difficult to forge or alter post-hoc. This is relevant for compliance, forensics, and incident response in production deployments.
- Kill switch: A mechanism exists to halt all model activity in emergency scenarios — a hard stop that doesn't depend on the model's own judgment at that moment.
The key architectural distinction is where the defense lives. External guardrails can be bypassed if an attacker convinces the model to format a dangerous output in a way the guardrail doesn't recognize. Because ViKernel is in the weights, it has visibility at the tool-call level regardless of how the upstream reasoning was manipulated.
VIDRAFT CEO Kim Minsik described the threat shift to JoongAng Ilbo as follows: attackers are moving from targeting code to targeting the AI models themselves. In a world where models don't just generate text but actually execute actions, the relevant security boundary is the action layer — and that boundary needs to hold even when the model has already been deceived.
Benchmarks & results
The public benchmark used the following setup:
- 20 AI models evaluated
- 8 simulated tools provided to each model
- 100 attack prompts issued per model
Key results:
- All 20 models scored F — the lowest possible rating — on ViKernel Scanner's A-to-F scale
- The median model attempted a dangerous action on 40 out of 100 checks, meaning half of tested models had a dangerous-action attempt rate above 40%
- Across the full test run, 897 dangerous tool calls were logged — and all 897 were blocked before execution by ViKernel, with zero bypasses recorded
The full per-model breakdown is publicly available. VIDRAFT has also announced a $4,500 prize hackathon in which participants are invited to attack ViKernel directly — an adversarial red-teaming event open to anyone.
For edge-deployed models specifically, the risk surface is notable: edge models commonly have access to local tools and local filesystems, amplifying the consequence of any single unsafe call.
How to try it
ViKernel Scanner is publicly available on Hugging Face at no cost, with source code released under the Apache 2.0 license. It is compatible with any OpenAI-compatible endpoint.
# Install via pip (check Hugging Face for the current package name) pip install vikernel-scanner # Or browse/download via the Hugging Face CLI huggingface-cli download VIDRAFT/vikernel-scanner ⚠️ The exact package name and repo path above are illustrative of the general access pattern described in the source. Verify the current canonical name on Hugging Face before running.
A generic OpenAI-compatible scan would point the scanner at any endpoint you control. The tool returns an A–F rating in roughly one minute. The 20-model benchmark results are publicly posted alongside the tool.
FAQ
Q: Why does ViKernel need to be in the weights rather than sitting in front of the model as a proxy?
A: Prompt-level and output-level filters operate on text. A sufficiently crafted adversarial prompt can maneuver a model into producing a dangerous tool call that looks benign to a text classifier. ViKernel intercepts at the tool-dispatch layer — after the model has decided to act, but before the action runs — which is a different and harder-to-bypass control point.
Q: Can I use ViKernel Scanner to test a model I'm already running behind an OpenAI-compatible API?
A: Yes. The scanner is designed to work with any OpenAI-compatible endpoint, so it can evaluate models you're hosting locally, on-premises, or via third-party inference providers without requiring changes to your serving stack.
Originally reported by 넷이즈 网易 (중국) (2026-11-19) — source article.