Trust Evidence

Every proof here has a source. See for yourself.

Cases, certifications, patents, and papers, gathered in one place. Certificate numbers check out with the issuing body, patents in the national registry, papers by their venue. Open any of them yourself.

0% Raw data egress to external AI · on-prem deployment
0.94 PII detection F1 automated substitution
Blocked → Enabled External LLM use on data previously off-limits
Case Records

Operational case records

Blocked workflows, unblocked.
What changed, and how to reproduce it for audit.

01 Insurance LLM-assisted claims processing: sensitive-data substitution
Before ExposedSensitive data in prompts

Claims documents containing policyholder names, ID numbers, and medical details were sent directly to an external LLM API. The compliance team blocked the workflow.

After SubstitutedSensitive fields PreservedOutput usability

LLM Capsule substituted sensitive fields with restorable stand-ins before submission. Outputs returned and were reconstructed locally for downstream system use.

Trust boundary: raw policyholder & medical data stays inside
What changed

The LLM Capsule layer was inserted into the workflow. Substitution covered names, IDs, dates, and medical field patterns. Sensitive raw values stayed in the local reconstruction mapping.

How it works
  1. 1 Detection names, IDs, dates, medical field patterns
  2. 2 Substitution restorable stand-ins before submission
  3. 3 Capsule transmission only the substituted capsule reaches the LLM
  4. 4 Reconstruction outputs reconstructed locally for downstream systems
Reproduce

Each substitution run is logged and bound to a fixed release state, so the workflow is reproducible with the same substitution logic for audit and regression verification.

Evidence
State Card Substitution Log Reconstruction Mapping Record Re-run Record
Before BlockedExternal LLM (before)

In a network-separated (air-gapped) environment, external LLMs could not be used at all, and concern over classified-context exposure blocked AI adoption entirely.

After EnabledExternal LLM (after) 0%Raw data egress

LLM Capsule substitutes the sensitive context locally. Only the substituted capsule, not the original, is processed by the external LLM, and the result is rehydrated locally inside the boundary. The raw classified context stays inside the boundary, aligned with N2SF guidelines.

Trust boundary: raw context stays inside the air-gapped boundary
What changed

Sensitive context is substituted with local stand-ins before processing and reconstructed locally after. No original data leaves the enclave.

How it works
  1. 1 Sensitive context classified operation detail
  2. 2 Substitution performed locally inside the enclave
  3. 3 Execution capsule processed, 0% raw data egress
  4. 4 Reconstruction rehydrated locally; N2SF guideline aligned
Reproduce

Every substitution/reconstruction event is logged locally, so any processed request can be reconstructed and inspected within the boundary.

Evidence
Local Reconstruction Mapping Audit Log N2SF Alignment
Before BlockedLLM on internal docs (before)

Internal documents (contracts, HR records, customer files holding personal and confidential information) could not be sent to ChatGPT, Claude, or Gemini for summarization or search. The security team blocked it.

After EnabledSensitive LLM use (after) LocalPII detect & restore

LLM Capsule detects and substitutes personal and confidential fields locally before the request reaches an external LLM, then restores the original values in the response, so staff can summarize and query internal documents without raw data leaving the organization.

Trust boundary: raw internal documents stay inside; only substituted values reach the external LLM vendor
What changed

The LLM Capsule layer was placed between internal tools and the external LLM. Detection covered names, contacts, IDs, and account/contract identifiers; raw values stayed in the local reconstruction mapping.

How it works
  1. 1 Detection names, contacts, IDs, account/contract identifiers
  2. 2 Substitution restorable stand-ins before the request leaves
  3. 3 Layer only the substituted request reaches ChatGPT / Claude / Gemini
  4. 4 Reconstruction original values rehydrated locally in the response
Reproduce

Each request's substitution is logged and bound to a fixed release state, so a given document-Q&A run can be reconstructed for audit.

Evidence
Substitution Log Reconstruction Mapping Record Audit Log
Before BlockedValidation status (before)

Real patient records required for model validation could not be accessed due to regulatory constraints. The validation pipeline stalled.

After UnblockedValidation status (after) DP-safePrivacy method

Differential-privacy-safe (DP-safe) replacement records matching real distribution characteristics, without containing real identifiable information, unblocked the validation pipeline.

Trust boundary: real patient identifiers stay out of the pipeline
What changed

Non-accessible real records were replaced with DP-safe equivalents. Data distribution preserved, compliance review passed, validation pipeline resumed without modification.

How it works
  1. 1 Restricted input real patient records (regulatory block)
  2. 2 Replacement DP-safe, distribution-matched equivalents
  3. 3 Privacy no real identifiable information in the dataset
  4. 4 Outcome validation unblocked; compliance review passed
Reproduce

The replacement dataset is versioned and bound to a fixed data state, so the validation run is reproducible with the same distribution on demand. Audit trail maintained throughout.

Evidence
State Card DP Audit Log Dataset Version
Before RestrictedRaw OT data (before)

OT/ICS network data carried sensitive operational details, so it could not be sent to an external AI for automated threat analysis.

After EnabledAI threat analysis (after) Structure-preservingSubstitution

Structure-preserving substitution lets an AI agent analyze the network data and answer threat questions. Sensitive values are replaced with stand-ins while relationships are preserved. (Integrated with a global OT security platform's detection solution.)

Trust boundary: sensitive operational values stay inside; only substituted values leave
What changed

Network-data sensitive fields are substituted while topology and relationships are kept intact, so the agent can reason over realistic context.

How it works
  1. 1 Restricted input OT/ICS network data with operational detail
  2. 2 Substitution structure-preserving (topology kept intact)
  3. 3 Analysis AI agent reasons over realistic, substituted context
  4. 4 Integration works with a global OT security detection platform
Reproduce

The substituted dataset and the agent's analysis are bound to a fixed data state, so the same analysis can be re-run and verified.

Evidence
Substituted Dataset Agent Analysis Log Structure Map
Before ManualRelease screening (before)

Users exporting sensitive aggregate statistics required manual, per-desk substitution and release review, which was inconsistent and hard to audit.

After AutomatedScreening (after) 0.94PII detection F1

A per-desk substitution module plus a multi-agent pipeline detects, traces, and substitutes personal information in aggregate data, automating and standardizing the release-review process.

Trust boundary: PII removed and traced before any release
What changed

Release State fingerprints the data before and after substitution, so which records were transformed, and how, is traceable for audit.

How it works
  1. 1 Before manual, per-desk substitution (hard to audit)
  2. 2 Detection multi-agent PII detection (F1 0.94)
  3. 3 Pipeline detect, trace, substitute, standardize
  4. 4 Audit release state fingerprinted before & after
Reproduce

A prior release can be replayed against its bound Release State, reproducing the screening process for regulatory inspection.

Evidence
Substitution Report Release Audit Log Detection Trace State Card
Certifications · Patents · Research

Independent validation

Certifications, patents, and peer-reviewed research. Each checkable at the source.

Certification
ISO 27001: Information Security Management
ISO/IEC 27001:2022 · Certified by Korea Foundation for Quality (KFQ) · 2026

International standard for information security management. Demonstrates a systematic approach to managing sensitive information.

Certification
ISO 42001: AI Management System
ISO/IEC 42001:2023 · Certified by Korea Foundation for Quality (KFQ) · 2026

International standard for AI management systems. Demonstrates responsible AI governance and risk management.

Certification · GS Grade 1
GS Certification (Grade 1): LLM Capsule
Korean SW Quality Certification, Grade 1 · LLM Capsule 2024 · listed on the public Innovation Procurement Marketplace

Top-grade (Grade 1) Korean government software-quality certification, supporting public-sector procurement via the Innovation Marketplace.

Government Track
KISA Fast Track (2024)
Korea Internet & Security Agency (KISA) · 2024

Selected for the KISA information-security industry Fast Track program.

View full details in the Trust Center
Patent · KR Registered · US Allowed
AI-Based Service Providing Method Without Leaking Private Information and Client Apparatus
KR Reg. No. 10-2757651 (App. 10-2023-0133086, Registered 2025-01-16) · US App. No. 18/908,054 (Filed 2024-10-07, allowed for registration 2026-07)

Core LLM Capsule patent. Method and client apparatus for AI services without exposing private information. Registered in Korea; allowed for registration in the US (2026-07).

Patent · Pending · KR + US
Method for Providing Security for On-Device Artificial Intelligence Models
KR App. No. 10-2025-0003223 (Filed 2025-01-09) / 10-2026-0000037 (priority, Filed 2026-01-02) · US App. (Ref. PO25-025-US)

Security provisioning method for AI models running on-device, supporting the on-prem / air-gapped execution path. Korean priority applications with a corresponding US filing.

Patent · Pending · KR · Expedited
Data Management Method and System for AI Execution Control
KR App. No. 10-2026-0053050 · Filed 2026-03-24 · Expedited examination granted 2026-04-08

Method and system for controlling and managing data state within AI execution environments. Expedited examination granted.

Patent · Pending · KR
Method and Inference Apparatus for Building Deep Learning Models Robust to Private Information Exposure
KR App. No. 10-2023-0074745 · Filed 2023-06-12

Deep-learning model construction robust to private-information exposure. Applicant: Ewha Womans University (co-research).

Patent · Pending · KR
Method and Analysis Apparatus for Building AI Models that Process Heterogeneous Datasets
KR App. No. 10-2023-0013029 · Filed 2023-01-31 · Under examination

AI model construction method for heterogeneous datasets. Applicant: Ewha Womans University (co-research).

Substitution & structure preservation
IEEE BIBM 2023
Privacy-Preserving Publishing of Individual-Level Medical Data for Cloud Services
Ho Bae, Heonseok Ha, Siwon Kim · IEEE BIBM · Istanbul, Dec 2023

Formalizes how individual-level medical records can be released to cloud services under strict privacy constraints while staying useful. It frames the exact problem LLM Capsule addresses: letting an organization use sensitive data with an external service without handing over the raw records.

IEEE/ACM TCBB 2022
DNA Privacy: Analyzing Malicious DNA Sequences Using Deep Neural Networks
Ho Bae, Seonwoo Min, Hyun-Soo Choi, Sungroh Yoon · IEEE/ACM Transactions on Computational Biology and Bioinformatics · 2022

A deep-learning approach to security and privacy on genomic sequences, among the most tightly regulated data classes there is. It shows how far the founding team's privacy research reaches, well beyond the document and tabular data an AI data layer handles day to day.

PSB 2020
AnomiGAN: Generative Adversarial Networks for Anonymizing Private Medical Data
Ho Bae, Dahuin Jung, Hyun-Soo Choi, Sungroh Yoon · Pacific Symposium on Biocomputing · 2020

Anonymizes sensitive medical records with a GAN while preserving the statistical structure downstream models rely on. LLM Capsule works from the same structure-preserving principle: it substitutes fields before they reach an external LLM and restores them afterward, so the data stays usable without being exposed.

Membership & inference-attack defense
BMVC 2022
MPGAN: Membership Privacy-Preserving GAN
Heonseok Ha, Uiwon Hwang, Jaehee Jang, Ho Bae, Sungroh Yoon · BMVC · London, Nov 2022

Trains generative models that resist membership-inference attacks, so an adversary can't tell whether a given record was in the training data. The same concern drives how LLM Capsule keeps individual records from being re-identifiable in what a model returns.

ACM AsiaCCS 2022
Membership Feature Disentanglement Network
Heonseok Ha, J. Jang, Y. Jeong, S. Yoon · ACM Asia Conference on Computer and Communications Security · 2022

Disentangles membership-revealing features from model representations, lowering exposure to inference attacks. It is the kind of privacy guarantee enterprise buyers ask about before sensitive data ever touches a model.

Distributed-learning security & foundations
ESORICS 2024
VFLIP: A Backdoor Defense for Vertical Federated Learning via Identification and Purification
Yungi Cho, Woorim Han, Miseon Yu, Younghan Lee, Ho Bae, Yunheung Paek · ESORICS · 2024

First backdoor defense specialized for Vertical Federated Learning. It identifies and purifies backdoor-triggered embeddings at inference.

ESORICS 2023
FLGuard: Byzantine-Robust Federated Learning via Ensemble of Contrastive Models
Younghan Lee, Yungi Cho, Woorim Han, Ho Bae, Yunheung Paek · ESORICS · 2023

Byzantine-robust federated learning that detects malicious clients via an ensemble of contrastive models, strong under non-IID data.

JMLR 2025
Regularizing Hard Examples Improves Adversarial Robustness
Hyungyu Lee, Saehyung Lee, Ho Bae, Sungroh Yoon · Journal of Machine Learning Research · 2025

Adversarial robustness method that regularizes hard examples to improve robust generalization.

Sensors 2024
Evaluation of Malware Classification Models for Heterogeneous Data
Ho Bae · Sensors (MDPI) · 2024

Study of malware-classifier explainability on heterogeneous data. Existing explanations fall short, and high accuracy can give a misleading sense of security.

Selected publications by the CUBIG founding team and collaborators. Full publication list and citation counts available on request.

Awards · Partners · Press

Market recognition

Awards, partnerships, and press, with dates and links.

Industry Award
Deutsche Telekom T-Challenge 2026: 2nd Place
T-Mobile / Deutsche Telekom · 2026

Placed 2nd in the global open-innovation program with substitution and local restoration technology.

Government Award
Information Security Product Innovation Award: Minister of Science and ICT Prize
Ministry of Science and ICT · 2024-11-26

Grand Prize, Information & Physical Security category.

Recognition
2026 Emerging AI+X Top 100
Korea AI Industry Association · 2026

Selected for the zero-access architecture and privacy technology.

Government Program
Selected Supplier: 2026 AI Voucher Program
Ministry of Science and ICT · NIPA · 2026

Selected as a supplier for the 2026 AI (Cloud) Voucher program, letting institutions adopt CUBIG solutions, including LLM Capsule, via government vouchers.

Government Program
Ultra-Gap Startup 1000+ (DIPS) · Global ICT Future Unicorn
Ministry of SMEs and Startups · KISED · 2025

Selected as a top deep-tech startup (AI / big data) in the Ultra-Gap Startup 1000+ project, and as a Global ICT Future Unicorn the same year.

Startup Recognition
NVIDIA Inception · SKT × Hana Bank AI Accelerator
2024–2025

NVIDIA Inception member and joint AI startup accelerator selection.

Recognition
Startup World Cup Finalist · NextRise Global Innovator
2024

Global startup competition finalist and innovator selection.

Marketplace
AWS Marketplace
LLM Capsule listed · procurement-ready

Available for direct purchase and private-offer deployment through AWS, simplifying enterprise procurement.

Infrastructure
NVIDIA Inception
Member program

Access to NVIDIA's technology and go-to-market resources supporting on-prem and accelerated deployment.

Cloud
NAVER Cloud
Deployment partner

Korean public/enterprise cloud deployment path, relevant for data-residency-constrained customers.

Press Coverage
CUBIG Validates LLM Capsule's Technical Efficacy Across Global Industrial Sites ↗
Korea Economic Daily Business · April 22, 2026

Coverage of LLM Capsule's context-preserving data layer architecture being validated in production alongside global industrial cybersecurity firm Claroty, with original data staying inside the customer environment.

Press Coverage
CUBIG Listed in Gartner Hyper-Synthetic Data Report: Only Asian Vendor Named ↗
Electronic Times (etnews) · July 30, 2025

Gartner's "Emerging Tech: Trends for Hyper-Synthetic Data" (June 2025) named CUBIG as one of 26 global key vendors following a 4-month deep-dive interview process and analysis of 60+ real-world use cases. The only Asia-based company on the list.

Press Coverage
LLM Capsule Emerges as Essential Tool for Safe Enterprise AI Adoption ↗
Korea Economic Daily · April 1, 2025

As enterprises and public institutions hesitate to deploy generative AI over data leakage concerns, LLM Capsule's real-time sensitive-data blocking for ChatGPT, Claude, and Gemini is gaining recognition as the go-to solution for compliant AI use.

Interview
[VivaTech 2025] CUBIG: "Designing the Future of AI Data Security" ↗
NewsPim · June 4, 2025

CEO Bae Ho interviewed at VivaTech 2025 in Paris, covering CUBIG's three core technologies (DTS, LLM Capsule, and DataXpert) and explaining why air-gapped deployment is driving rapid uptake in finance, healthcare, and defense.

Press Coverage
LLM Capsule Selected for South Korea's Government Tech Pilot Purchase Program ↗
TechWorld News · June 2, 2025

The Ministry of SMEs and Startups selected LLM Capsule for its technology pilot procurement program, clearing rigorous evaluation on technical capability, market viability, and innovation, opening a direct public-sector deployment path.

Press Coverage
CUBIG to Unveil Next-Generation Synthetic Data Technology at World IT Show 2025 ↗
AVING News · April 22, 2025

CUBIG showcased LLM Capsule at WIS 2025 at COEX Seoul, highlighting standalone air-gapped operation, beta testing with global SaaS and healthcare companies, and planned cloud integration with AWS and Google Cloud.

FAQ

Common questions

Context-preserving AI keeps sensitive data inside your environment. The AI works only on structure-preserving substitute values, and results are reconstructed locally into the workflow you already run. So they stay usable and verifiable.
Operational evidence is concrete, verifiable documentation that an AI system behaves as claimed in real production conditions: what was blocked before, what changed after deployment, the deployment reference, and the certifications, patents, and peer-reviewed research that back it. It is the difference between a marketing claim and a record a procurement or audit team can verify.
Through third-party certifications (ISO/IEC 27001, ISO/IEC 42001, KISA Information Security Fast Track, GS Certification), registered patents covering privacy-preserving AI services and on-device AI security, peer-reviewed publications by the founding team, and deployment references across regulated industries.
Certifications show an independent body has audited the security and AI-management practices; patents show the method is novel and documented; peer-reviewed research shows the underlying techniques withstand academic scrutiny. Together they let a buyer rely on more than the vendor's own word.
No. Across every record on this page, only structure-preserving substitute values are shared with AI. Sensitive values stay inside the environment through structure-preserving, differential-privacy-based substitution, and outputs are restored locally into the originating workflow.
The certifications, patents, research papers, and press items on this page each cite a primary source you can check independently, without taking our word for it. The certifications have issuing-body certificate numbers confirmable with the certifier; the patents list their official KIPO numbers, lookupable in the national registry or Google Patents; the papers cite venue and authors traceable through DBLP; and every press item links to its original article.
Credible proof is evidence a third party can verify without relying on the vendor's word, not a testimonial or a marketing line. For an AI data layer that means several signals together: operational records of what was blocked and what changed, independent certifications, registered patents in a public registry, and peer-reviewed publications by the founding team. Any one alone is weak; together they let a buyer rely on more than a promise.

Now see it on your own data.

One workflow blocked by data exposure is enough to start. We will show the same evidence on your data, your systems, and your constraints.

See the Architecture Run a sample proof