Skip to main content
Korean financial and public-sector AI security guidelines

Financial and public-sector security
requirements, built into the AI gateway

Syncgate provides the input and output filtering, personal data protection, log retention and access control that Korean AI security guidelines for financial institutions and public institutions require. Each control is shown with the check items it addresses.

Guideline mapping

Six controls
and the check items they address

Personal (credit) information detection and pseudonymization

Detects 13 data types, including resident registration numbers, account numbers, card numbers and driver's license numbers, in input and output, and blocks or masks them. Only pseudonyms are sent to the model, and the original values are restored when the answer is shown to the user. It can also warn at the input stage and ask the user to record a reason.FSC guideline check 7 · FSS AI RMF privacy · NIS guidebook M13

Layered defense against prompt injection and jailbreaks

On top of blocked-word and regular-expression rules, a guardrail model that judges each sentence in its context runs a second check. With “Pause after repeated violations” turned on, a user is warned after 5 violations and blocked for 30 minutes after 10. These defaults match the compliance example in the guidelines.FSC guideline checks 6, 9 · NIS guidebook M13, M15

Identifying security classification levels (C · S · O)

Classifies documents as Classified, Sensitive or Open (C · S · O) and allows, labels or blocks each level. Classified documents are left out of results sent to external AI tools by default.N2SF Model 2 “AI integration system” · IF-14

Log retention and tamper evidence

Records requests, guardrail decisions and administrator activity. The audit log is kept indefinitely by default, checked for tampering with a hash chain, and sent to your SIEM over syslog in real time.National Cybersecurity Basic Guidelines Art. 55 · NIS guidebook M09 · Compensating controls for financial network separation

Approved models only

Users and company systems do not call models directly. Their requests are sent through the gateway. Only providers and models an administrator has approved are allowed, per group or per user, and request and token limits are set for each user. On an air-gapped network, only self-hosted models are used.NIS guidebook M17 · FSC guideline check 12 · N2SF Model 2

MFA, least privilege and human approval

SSO, SCIM and two-step verification with TOTP or passkeys are supported, and least privilege is applied per group. Two-step verification can be made mandatory for administrators. Tools that change external systems run only after human approval, and the chat input shows the security warning your organization sets.FSC guideline check 24 · National Cybersecurity Basic Guidelines Art. 76 · NIS guidebook M05 · M30 · A-M12

Authentication, permissions and audit

Ready for your
security review

Authentication, auditing, encryption and deployment options: the items security teams check first.

Single sign-on and automatic provisioning

Supports sign-in with Microsoft Entra ID, Google Workspace, SAML 2.0, OpenID Connect or LDAP, SCIM user sync, and two-step verification (authenticator app, passkey or security key).

Sealed audit logs

Records time, user, action, resource and IP. Each record carries the hash of the one before it, so integrity can be verified.

Encryption key management

Use the built-in key or integrate with Azure Key Vault, AWS KMS, Google Cloud KMS or Vault. Key rotation is supported.

Per-organization databases

In an installation used by several organizations, each organization has its own database and a dedicated role, so their data is kept separate.

Retention periods and legal hold

Set a retention period for each type of record, and records are cleaned up automatically when it ends. Accounts and conversations under legal hold are excluded from cleanup.

Granting and revoking support access

The service operator can sign in for support only while an administrator allows it. The access period can be set from 4 hours to 30 days, or to “Until revoked,” and the administrator can end the access at any time.

Data location

Data stored internally
and the scope of external communication

Syncgate is installed in an environment your company designates. External communication is limited to calls to models an administrator has approved and to communication with the sources and tools an administrator has connected.

Company network · private cloud
Syncgate serverPermission check, guardrails, model gateway and logging. Installed with Docker Compose or Helm.
DatabasesStores conversations, settings, audit logs and indexes in PostgreSQL. With several organizations, each organization gets a separate database.
File storageStores uploaded files and result files. Server storage is the default; S3-compatible storage can also be used.
Harness runtimeAn isolated container that reads files and runs code.
In-house models (optional)Ollama, vLLM, LM Studio, NVIDIA NIM and others. On an air-gapped network, only in-house models are used.
Internal document sources and toolsInternal wikis, storage, databases and internal APIs.
Internal key management service and SIEMKey management services such as Vault or KMS, and a SIEM that receives syslog.
External services (only those an administrator approves)
Approved model providersRequests go only to models an administrator has connected and allowed. Only content the guardrails have checked is sent.
Cloud document sourcesCloud services an administrator has connected with a connector.
Web search (if enabled)Used only when the web search feature is enabled.

Air-gapped network

A configuration with no connection to external services. With only in-house models connected, normal operation has been verified from installation through to conversations.

License

The license file is checked on the server itself, with no connection to an external server.

Service operator access

The service operator can sign in only during the period an administrator sets, and that access is logged as well.

Self-hosted and SaaS compared

Operating conditions that depend
on where it is installed

The differences between a typical SaaS AI service and self-hosted Syncgate. This does not refer to any specific company's product.

ItemTypical SaaS AI serviceSelf-hosted Syncgate
Where data is storedThe vendor's cloudPostgreSQL and file storage on your servers or private cloud
Air-gapped networkInternet connection requiredInstallation, indexing, search and conversations verified without an internet connection
Available modelsModels the service choosesModels an administrator approves, from 23 providers and in-house models
LogsScope and period set by the serviceSealed audit log, retention period per record type, and forwarding to a SIEM over syslog
Encryption keysManaged by the vendorBuilt-in key, or your company's Azure Key Vault, AWS KMS, Google Cloud KMS or Vault
UpdatesWhen the vendor decidesRuns at a time the administrator sets (can be turned off or run immediately); the previous version is restored automatically if an update fails
Time to deployImmediately after sign-upRequires server preparation and installation
Server operationsThe vendorYour company (operator console and manual provided)
FAQ

Questions security teams often ask

Are our conversations used to train models?

Syncgate has no feature that trains models on conversations. How an external provider's models handle data is governed by your contract with that provider, so check the contract terms when you approve a provider. On an air-gapped network, only in-house models are used.

Can audit logs be deleted or altered?

Deleting or editing an audit log shows up in the integrity check, because each record contains the hash of the one before it. When logs are cleaned up after their retention period, the cleanup itself is recorded, so integrity can still be checked.

How are administrator accounts protected against takeover?

Two-step verification can be made mandatory for administrators, with an authenticator app, passkey or security key. SSO sign-in can also be enabled.

What happens if an employee enters customer information into the AI?

Guardrails detect 13 types of personal data in input and output and block, mask or pseudonymize them. They can also be configured to warn users at the input stage and require them to record a reason.

Can we get the documents we need for our review?

Yes. Send a request through Request security documentation and our team will contact you.

Request the documents
your security review needs

We explain the installation setup, the guideline items addressed, and the log and retention policies, as your review requires.

Product videosReal product screens · Example company ACME Logistics · 16 videos