Trust & Security
Not trained on · Not logged · Not retained
No data sent through Repattern AI is ever used to train AI models.
Beyond training, we have disabled API logging at every LLM provider we use.
Your conversations are left on LLM servers in no form at all.On top of that, Repattern AI runs its own Prompt Guard, which
automatically masks sensitive data and blocks banned terms before they leave.
At a glance
| Item | Status |
|---|---|
| Data used for AI model training | Not used |
OpenAI API store: false | Applied - forces request data not to be stored on OpenAI servers |
| LLM provider API logging | Disabled everywhere |
| Zero Data Retention (ZDR) | Application in progress - excludes even abuse-monitoring logs |
| Prompt Guard (in-house security layer) | Sensitive-data masking · banned-term blocking |
| Encryption at rest | AES-256 |
| Encryption in transit | TLS 1.3 |
| Infrastructure | Seoul region cloud |
| Suspected leak | Investigation · filing with the provider on your behalf |
Is what our employees type into the AI used for training?
No. It is never used for training.
Repattern AI uses the paid APIs of the major LLMs - OpenAI, Anthropic (Claude), and Google (Gemini).
All of these providers share the same baseline policy: data sent through a paid API is not used for model training.
OpenAI (GPT)
"As of March 1, 2023, data sent to the OpenAI API is not used to train or improve OpenAI models (unless you explicitly opt in to share data with us)."
- Source: OpenAI API Data Policies +1
- Participating in model training requires an explicit opt-in in the API dashboard
- Where to check: Organization Data Controls > Sharing
- Repattern AI keeps this setting Disabled
- Code-level protection: Repattern AI sends
store: falseon OpenAI API calls, forcing request data not to be stored on OpenAI servers (applied automatically to every OpenAI call)
Anthropic (Claude)
"By default, we will not use your inputs or outputs from our commercial products (e.g. Claude for Work, Anthropic API, Claude Gov, etc.) to train our models."
- Source: Is my data used for model training?
- Data can be used for training only when a user explicitly opts in through the feedback button
- Where to check: Workspaces > Privacy Settings
- Repattern AI keeps the user feedback feature (Rate chats) Disabled
Google (Gemini)
"Google doesn't use your prompts (including associated system instructions, cached content, and files such as images, videos, or documents) or responses to improve our products."
- Source: Gemini API Additional Terms of Service - How Google Uses Your Data (Paid Services)
- On paid tiers, prompts (including system instructions, cached content, and files) and responses are not used to improve Google products
Even if it is not used for training, is our conversation recorded somewhere?
Good question. Training and logging are two different things.
Many people hear "we don't train on it" and stop worrying - but if logging is enabled, conversations can still sit on the provider's servers for a retention period.
| Aspect | Training | Logging |
|---|---|---|
| What it is | Using data to improve the AI model | Storing API call records on the provider's servers |
| Purpose | Better model performance | Debugging, analytics, monitoring |
| Risk | Data may be absorbed into the model | Conversations may remain on provider servers |
Repattern AI goes one step further and disables logging at every LLM provider.
What matters here is that most LLM providers have API logging on by default. Use the API without taking action and your conversations can be stored on provider servers for a retention period. Google (Gemini) is the only one whose paid API has logging off by default.
Knowing this, our service operations team turned logging off by hand in each provider's admin console.
Logging settings by provider
| LLM provider | Default | Repattern AI setting | Where it is set |
|---|---|---|---|
| ON by default | Turned OFF | Organization > Data Controls > Data Retention | |
| ON by default | Turned OFF | Workspaces > Privacy | |
| OFF by default | Kept OFF | AI Studio > Logs |
Use these APIs without taking action and conversations can be recorded - but because Repattern AI disabled logging at every provider, your conversations are never even written down on LLM servers.
But some data can remain even with logging off?
Correct. After training and logging, a third concept enters - abuse monitoring.
To prevent misuse of their services, LLM providers keep abuse-monitoring logs for a period of time, separately from ordinary logging. This is not a copy of your whole conversation; it covers only the very small number of requests suspected of violating abuse policy.
| Aspect | Training | Logging | Abuse monitoring |
|---|---|---|---|
| What it is | Used to improve the AI model | API call records stored | Surveillance to prevent misuse |
| Scope | All data | All data | Only requests suspected of policy violations |
| Repattern AI action | Not used | All OFF | ZDR application in progress |
What is abuse monitoring?
Abuse monitoring is a minimal safeguard for detecting and preventing misuse of the service. It catches cases such as:
- Attempts to generate illegal content: child sexual abuse material, terrorism-related content, and other legally prohibited content
- Building cyber-attack tooling: malware, phishing emails, hacking tools
- Large-scale automated abuse: spam generation, mass disinformation, automated fraud
- Terms-of-service violations: serious breaches of each provider's usage policies
Ordinary work conversations are not what abuse monitoring is for. Normal questions, drafting documents, writing code, translating - none of it trips this monitoring.
Abuse-monitoring retention by provider
Even with logging off, abuse-monitoring logs may be retained for the periods below:
| LLM provider | Default abuse-monitoring retention | Notes |
|---|---|---|
| Up to 30 days | Only requests suspected of policy violations | |
| Up to 30 days (up to 2 years for confirmed violations) | Deleted automatically; only confirmed violations are kept long-term | |
| Up to 55 days | Used solely for policy enforcement |
You can rest easy here. These logs are not a copy of your whole conversation - they hold only the very small number of requests that seriously violate abuse policy, and only briefly. Normal work conversations are not covered, and the records are deleted automatically once the retention period passes.
Zero Data Retention (ZDR) - excluding even abuse-monitoring logs
Repattern AI goes one step further still.
Abuse-monitoring logs do not cover ordinary work conversations, but to rule out even that sliver of retention, Repattern AI is applying for Zero Data Retention (ZDR) with each provider.
What is ZDR?
Zero Data Retention is a policy guaranteeing that data sent through the API remains on the provider's servers in no form at all once the response has been generated. It sits one level above disabling logging.
1graph LR 2 A["No training\nprovider default\n✓ applied"] --> B["Logging OFF\nset by flow\n✓ applied"] 3 B --> C["store:false\nenforced in code\n✓ applied"] 4 C --> D["Abuse monitoring\n30-55 days by default\n✓ applied"] 5 D --> E["ZDR\n0 days\n⏳ in progress"] 6 style A fill:#d4edda,stroke:#28a745,color:#155724 7 style B fill:#d4edda,stroke:#28a745,color:#155724 8 style C fill:#d4edda,stroke:#28a745,color:#155724 9 style D fill:#d4edda,stroke:#28a745,color:#15572410 style E fill:#fff3cd,stroke:#ffc107,color:#856404| Aspect | Logging OFF (in effect today) | ZDR (in progress) |
|---|---|---|
| Used for training | No | No |
| API call logging | No | No |
OpenAI store: false | Applied (enforced in code) | Applied |
| Abuse-monitoring logs | Retained by default (30-55 days) | Not retained (0 days) |
| How it is applied | Dashboard setting + code setting | Separate contract/approval required |
ZDR application status by provider
| LLM provider | ZDR status | How to apply | Official documentation |
|---|---|---|---|
| Approval request sent | Contact sales → dashboard setting after approval | Data Controls Guide | |
| Approval request sent | Contact sales → sign a ZDR agreement | Zero Data Retention | |
| Approval request sent | GCP Support → per-project approval | Gemini API ZDR |
Once ZDR is in effect, abuse-monitoring logs included, your data will not remain on provider servers even temporarily. Repattern AI has sent ZDR approval requests to each provider and will update this page as they are granted.
Should we just trust the promise not to train on our data?
No - and Repattern AI does not rely on that promise alone.
The LLM providers state clearly that they do not train on this data, but Repattern AI took it one step further. In case something ever goes wrong, we built our own security layer that stops sensitive data from reaching the LLM in the first place.
That layer is Prompt Guard.
What is Prompt Guard?
Prompt Guard is a security filter that runs immediately before a user's message is sent to an external AI model. Independently of any provider policy, Repattern AI detects and blocks sensitive data itself.
The same masking and banned-term checks apply not only to typed text but also to the contents of uploaded files (PDF, Excel, images, and so on). Sensitive data inside a file is detected and protected before it reaches the LLM.
1graph LR2 A["User input\ntext + files"] --> B["Prompt Guard\n✓ sensitive-data masking\n✓ banned-term blocking\n✓ protected-term masking"]3 B --> C["LLM API"]4 style B fill:#e8f4fd,stroke:#1a73e8,color:#0d47a1Even with a provider's promise not to train on it, if the original data never leaves in the first place, the worry has nothing to attach to.
What exactly is protected?
Tier 1: Automatic personal-data protection (DLP)
Around 30 kinds of sensitive data are detected automatically, in Korean and English text alike. Detected values are anonymized before reaching the LLM and restored to their original values when the response comes back. Users do not have to think about any of it.
Coverage falls into two groups. Country-independent: payment card numbers, IBAN bank accounts (89 country formats), passport numbers, email addresses, IP and MAC addresses, GPS coordinates, API keys, private keys and credential-like values, and crypto wallet addresses. Korea-specific: resident registration, alien registration, driver's licence, business and corporate registration numbers, domestic bank accounts, addresses and phone-number formats. National identifiers outside Korea - a US Social Security number, for instance - are not matched by a dedicated pattern; they are covered only by AI context detection below, or by protected terms an administrator registers (Tier 2).
| Protection level | What is covered |
|---|---|
| Mandatory | Bank account number, passport number, and Korean national identifiers (resident registration, alien registration, driver's licence, business registration) |
| High | Phone number, email address, credit card number, address |
| Medium | IP address, MAC address and other technical identifiers |
| AI-detected | Names, organizations, locations (determined automatically from context by AI) |
Tier 2: Admin-defined protected terms
Every company protects different things. Administrators can register their own confidential terminology, project code names, internal system names and so on; any message containing those terms is masked automatically before it is sent to the LLM.
Tier 3: Banned terms - blocked from leaving at all
Some information must never leave, and masking is not enough. When an administrator configures banned terms, a message containing one is blocked immediately and never reaches the LLM at all.
| Protection level | What is covered |
|---|---|
| Baseline | Phone number, email address, bank account number, passport number, and Korean national identifiers (resident registration, alien registration, driver's licence, business registration) |
| Optional | Credit card number, address, IP address, MAC address and other technical identifiers |
| Advanced | Names, organizations, locations (determined automatically from context by on-premise AI) |
So where is our data stored? Is it safe?
All data is processed on our own servers and stored encrypted. Only your own organization can reach it.
Customer data is processed and stored on Repattern AI's own servers, not on LLM provider servers. Encryption is applied both at rest and in transit.
| Layer | Method | Details |
|---|---|---|
| Encryption at rest | AES-256 | Keys managed by the cloud managed key service |
| Encryption in transit | TLS 1.3 | Latest encryption protocol |
Handling by data type
| Data type | How it is handled |
|---|---|
| Conversation history | Stored fully encrypted |
| Document attachments | Kept until the chat room containing the file is deleted |
| Image files | Kept until the chat room containing the image is deleted |
Attachments and images share the lifecycle of their chat room. Delete the room and the associated files are deleted with it.
- Infrastructure: hosted in a Seoul region cloud
- The Repattern AI database is fully separated from LLM provider servers and is operated to the same security standard as the flow collaboration tool
- Database connections are also encrypted with TLS 1.3, and access is isolated per organization
Can the AI providers themselves be trusted?
Yes. The LLM providers Repattern AI uses hold global security certifications.
| Certification | |||
|---|---|---|---|
| SOC 2 Type 2 | O | O | O |
| SOC 3 | O | - | O |
| ISO 27001 | O | O | O |
| ISO 27017 / 27018 | O | - | O |
| ISO 42001 | - | O | - |
| CSA STAR | O | O | O |
| HIPAA | O | O | O |
| NIST 800-171 | - | O | - |
| FedRAMP | O | O (High) | O |
You can verify each provider's certifications directly at their official trust portals.
How does the data flow, end to end?
1flowchart LR2 User(["User"]) -->|"question\nTLS 1.3"| Server["Repattern AI server"]3 Server --> Guard["Prompt Guard\nsensitive-data masking\nbanned-term blocking\nprotected-term handling"]4 Guard -->|"sanitized prompt\nTLS 1.3"| LLM["LLM API\ntraining ✗ · logging ✗\nabuse logs: ✗ with ZDR"]5 LLM -->|"response\nmasking restored"| Server6 Server -->|"AES-256\nencrypted at rest"| DB[("chat history")]7 style Guard fill:#e8f4fd,stroke:#1a73e8,color:#0d47a18 style LLM fill:#fce8e6,stroke:#d93025,color:#7f00009 style DB fill:#e6f4ea,stroke:#34a853,color:#137333- User → Repattern AI server: question sent (TLS 1.3 encrypted)
- Prompt Guard: sensitive-data masking, banned-term blocking, protected-term handling
- Repattern AI server → LLM API: sanitized prompt sent (TLS 1.3 encrypted)
- LLM API → Repattern AI server: response received → masked values restored to their originals
- Repattern AI server: chat history stored (AES-256 encrypted, reachable only by your organization)
- LLM provider: does not use the data for training and does not log it (with ZDR in effect, abuse-monitoring logs are not retained either)
What do we do if we suspect a data leak?
The Repattern AI operations team supports you actively, from investigation through filing with the provider on your behalf.
Realistically, a data leak at an LLM provider is very hard to prove. Read the other way around, that also means each provider manages this so tightly that an actual incident would be global front-page news.
Even so, as the party providing the service, Repattern AI is ready to respond actively even in the unlikely event.
Response process for a suspected leak
1flowchart LR2 A["Suspicious signal found"] --> B["Report to the\nRepattern AI ops team"]3 B --> C["Internal investigation\nlog analysis\ntimeline reconstruction\nguard records cross-check"]4 C --> D["File with the provider on your behalf\nshare the outcome\nsupport further action"]5 style A fill:#fce8e6,stroke:#d93025,color:#7f00006 style D fill:#e6f4ea,stroke:#34a853,color:#137333- Intake: tell the Repattern AI operations team about the circumstances that look like a leak
- Internal investigation: we analyze Repattern AI-side security logs, Prompt Guard masking records, and the API call timeline
- Filing on your behalf: based on the findings, we file the official report on your behalf with the LLM provider concerned
- Follow-up: we share the provider's response and the outcome, and support further action if needed
As the service that serves your data, Repattern AI commits to supporting you actively, on your side, throughout any dispute with a provider.
Worth recording when you report
To let us investigate a suspected leak faster and more precisely, having the following ready helps a great deal.
| Item | Description | Tip |
|---|---|---|
| Time of discovery | When you first noticed the circumstances | As precise a date and time as possible |
| Screenshot | A capture of the screen showing the circumstances | Capture the full screen so the browser URL is visible |
| URL | The address of the page or service involved | Copy the full URL from the address bar |
| Conversation | The Repattern AI conversation where the information was entered | Room name and message timestamp |
| Suspected leak path | Where you found the information | External service, search results, a third party, etc. |
| Type of sensitive data | What kind of information appears to have leaked | Personal data, company confidential, project name, etc. |
Tip: the moment you spot something suspicious, take a screenshot immediately. Web pages can change or disappear at any time, so capturing the screen right away matters most.
Official security reporting channels of the LLM providers
Each provider runs an official channel for reporting data security issues. The Repattern AI operations team will file on your behalf, but the official channels are listed here for reference.
| Provider | Reporting channel | Notes |
|---|---|---|
| Bug Bounty Program (Bugcrowd) | Security vulnerabilities and data-related issues | |
| Responsible Disclosure (HackerOne) | Security vulnerabilities and privacy reports | |
| Google Bug Hunters | Security reports across all Google products |
You do not need to file anything yourself. Contact the Repattern AI operations team and we will assemble the necessary information, file with the provider on your behalf, and keep you posted on progress.
Summary
Using AI through Repattern AI is safe.
- LLM providers do not use data sent through the API for model training (baseline policy)
- Repattern AI goes further and has disabled API logging entirely
- We have applied for ZDR with every provider to rule out even abuse-monitoring logs (30-55 days)
- Prompt Guard blocks sensitive data up front, so the original data never reaches the LLM
- All data is stored encrypted with AES-256 and transmitted over TLS 1.3
- The LLM providers hold global security certifications including SOC 2 and ISO 27001
- In the unlikely event of a suspected leak, the Repattern AI operations team supports you actively, from investigation through filing with the provider
Your conversations are not trained on, not recorded, and kept safe.
And if something ever does happen, Repattern AI stays with you to the end.
Prompt Guard technical detail
Prompt Guard processing flow
1flowchart TD 2 A["User input\ntext + files"] --> B["Tier 3 · banned-term check"] 3 B -->|"if present"| C["🚫 Send refused\nnever reaches the LLM"] 4 B -->|"passes"| D["Tier 2 · admin protected-term masking\nconfidential terminology substituted"] 5 D --> E["Tier 1 · automatic DLP masking\n30+ PII types detected and substituted"] 6 E -->|"TLS 1.3"| F["LLM API"] 7 F --> G["LLM response received"] 8 G --> H["masking restored"] 9 H --> I["delivered to the user"]10 style C fill:#fce8e6,stroke:#d93025,color:#7f000011 style F fill:#e8f4fd,stroke:#1a73e8,color:#0d47a1DLP detection levels
| Level | Covered |
|---|---|
| CRITICAL | Credit card, bank account number, passport, API keys, encryption keys, and Korean national identifiers (resident registration, alien registration, driver's licence, business registration) |
| HIGH | Phone number, email address, address, GPS coordinates, employee ID, insurance policy number |
| MEDIUM | IP address |
| AI-detected | Names, organizations, locations (context-based) |
Scope
The same security masking applies to the file types below, not only to typed text.
- PDF files
- Excel files
- Images (after text extraction)
- Assistant system prompts
Learn more
You can review Repattern AI's privacy policies and security measures item by item.
| Document | Description |
|---|---|
| Self-assessment checklist (16 items) | AI privacy self-assessment results - 8 categories, 16 detailed items |
| Privacy impact assessment | PIA on the KISA template - risk identification and safeguards in place |
| Disclosure, processing, cross-border transfer | Where data goes and what protections apply |
| Security and safeguards | Encryption, access control, audit logs, destruction policy |