VIDRAFT's ByiKernel: A Security OS Embedded Directly Inside AI Model Weights
TL;DR: VIDRAFT (비드래프트) has developed ByiKernel, a security operating system that lives inside an AI model's weights rather than wrapping the model externally. It enforces data-exfiltration prevention, PII masking, anomaly-triggered kill-switches, and tamper-evident audit logs — all without a separate security layer outside the model. Because the security functions travel with the weights, they remain active even when the model is moved to a different server.
What it is
ByiKernel is a security OS that VIDRAFT has implemented within the weight parameters of an AI model itself. Traditional AI security approaches sit outside the model — inspecting inputs, outputs, and API calls via proxy layers, firewalls, or guardrail wrappers. ByiKernel takes a different architectural stance: the security logic is fused with the weights, so it migrates with the model and does not depend on the host infrastructure to enforce policy.
Specifically, ByiKernel is designed to handle four core security functions inside the model:
- Data-exfiltration blocking — unauthorized outbound transmissions are intercepted before they leave the agent's execution context.
- PII redaction — personally identifiable information (e.g., national ID numbers, bank account numbers) is masked before any external handoff.
- Kill-switch — anomalous behavior triggers a hard halt of model execution; resuming operation requires an operator-signed authentication key.
- Tamper-evident audit log — each attempted and executed action is recorded and hash-chained, so any retroactive modification of a log entry is detectable during verification.
VIDRAFT states the technique was applied to their own 27-billion-parameter-class model as well as to multiple external model families including Google Gemma and Alibaba Qwen.
How it works
At a conceptual level, VIDRAFT describes constructing computational and memory-like circuits within the model's internal representations — effectively encoding deterministic rule-evaluation logic into the weight space rather than appending it as an external post-processing step.
The security rule engine evaluates each action an AI agent intends to execute against a predefined policy before the action is carried out. This means the policy check happens inside the same inference pass, not as a downstream filter.
The audit trail uses cryptographic hashing to chain log entries sequentially. Because each record is linked to the hash of the previous one, tampering with an earlier entry breaks the chain and is flagged during integrity verification — a design pattern familiar to anyone who has worked with blockchain-style append-only logs.
VIDRAFT also reports running high-precision arithmetic experiments (multiplying two 300-digit numbers to produce a 600-digit result) as a proof-of-concept that sufficiently complex deterministic computation can be reliably embedded and executed within model weights. The company cites this as a foundation for the rule-evaluation and logging mechanisms.
Important caveat: The public demo replays pre-recorded responses generated by the security OS rather than running live inference on every interaction. The demo environment is not a real-time validation of the full in-weight security pipeline.
Benchmarks & results
VIDRAFT reported one quantitative result from internal testing:
- 661 test runs across 4 open-source AI models, using simulated tools and synthetic data.
- 0 unauthorized external transmissions detected across all 661 runs.
The company itself acknowledges these are internal test figures. Independent validation against the range of adversarial conditions present in real enterprise environments has not yet been published. Developers and security engineers should treat this as early-stage empirical data pending third-party audit.
How to try it
VIDRAFT has published a public demo on Hugging Face called "ByiKernel Desktop". The interface resembles a Windows-style desktop environment where you can interactively explore:
- Kill-switch activation
- Exfiltration blocking behavior
- PII masking
- Audit log verification
You can find it by searching for VIDRAFT's organization page on huggingface.co. No installation or API key is required for the Hugging Face Space demo.
No public GitHub repository, pip-installable package, or OpenAI-compatible API endpoint for ByiKernel has been announced at the time of this writing. If you want to evaluate production integration, direct contact with VIDRAFT is the current path.
FAQ
Q: Why embed security in weights rather than using an external guardrail layer?
A: External guardrails are tied to the deployment infrastructure — if the model is copied to a different server or cloud environment, those external controls don't automatically follow. Weight-embedded security travels with the model artifact itself, theoretically maintaining policy enforcement regardless of where the weights are hosted.
Q: Does having an audit log inside the model mean it automatically satisfies EU AI Act logging requirements?
A: Not necessarily. The EU AI Act requires high-risk AI systems to support automatic logging of operation, but regulatory compliance involves documentation, governance processes, and third-party conformity assessments beyond a technical logging capability alone. The audit log is a relevant building block, but compliance cannot be assumed from this feature alone.
Q: Which model families have been tested with ByiKernel?
A: VIDRAFT publicly stated the technology was applied to their own 27-billion-parameter-class model, Google Gemma, and Alibaba Qwen. No further model-specific details were disclosed in the source reporting.
Q: Is the current Hugging Face demo running live in-weight inference?
A: No. Per VIDRAFT's own disclosure, the public demo replays pre-recorded outputs. It is useful for exploring the UX and understanding intended behavior, but it does not constitute a live end-to-end demonstration of the in-weight security execution pipeline.
Originally reported by ZDNet코리아 (2026-11-03) — source article.
