Skip to content

Koraku RealtyReal estate & privacy-first AI

Contact

Security and privacy

Intended controls, clearly stated.

PrivaGate AI is in development. This page separates three things: what the product is designed to do, what this website actually does and has been tested for, and what we are not claiming. None of it is a certification, an SLA, or a guarantee.

1 · Product design intentions · Planned

What PrivaGate AI is designed to enforce.

These are controls the product is being designed around. They describe intent for a product that is not released. Production deployment requires its own configuration review.

  1. Authorization modelUser identity, role, and document access are confirmed before any retrieval runs.
  2. Tenant and document boundariesEach tenant has its own document store, mapping table, and audit log. The gateway does not cross tenant boundaries.
  3. Data minimizationThe payload to the cloud is the smallest context that can answer the task. The policy is per task, not global.
  4. Local mapping controlsOriginals, tokens, and restoration keys live in the customer environment. The mapping table is not exposed to Claude.
  5. Disclosure gatingA configured policy decides whether the request is approved for cloud, kept local, or escalated for review.
  6. Output validationResponses are checked for structure, source support, and disclosure risks before any token restoration.
  7. Audit-log minimizationAudit records capture the minimum event needed for accountability. Full confidential content is not duplicated into logs.
  8. Human oversightA licensed professional confirms the output. The system never auto-issues a regulatory document.

2 · This website · Implemented and tested

What this website does today.

These behaviors are implemented in the website’s theme and plugin and checked by the project’s automated tests in a local environment. This is not an independent audit.

  • No third-party requestsPages load their fonts, styles, and scripts from this site only. There are no trackers and no remote scripts.
  • Inquiry form protectionsHoneypot field, a per-form token, a limit of five submissions per address per ten minutes, and server-side validation of every field.
  • Inquiry data handlingThe message body is passed to email and is not written to the database. A small log keeps only metadata such as time, category, and message length.
  • Enumeration and legacy endpointsAuthor archives and the public user list are blocked, and XML-RPC is refused.
  • Response headersStandard hardening headers are sent. A Content Security Policy with script nonces is configured in report-only mode; it is not yet enforced.

3 · Claims not being made

What we do not promise.

  • We do not claim 100% anonymous, zero exposure, or zero hallucination. These are outcomes, not controls.
  • We do not claim APPI certification. Token replacement is not, by itself, legally meaningful anonymization.
  • We do not claim ISO or SOC 2 certification. Production deployments must be reviewed independently.
  • We do not claim Zero Data Retention approval. Actual data-handling terms are governed by the contract with the cloud provider.
  • We do not claim partnership with Anthropic. Claude is referenced as the planned cloud reasoning engine for approved context.
  • Removing names is not enough to eliminate identification risk. Context, token shape, and other fields can re-identify a person or transaction.

Security contact

Report a security concern.

Use the contact form with category “Security”. For sensitive disclosures, request a private channel by selecting “Request encrypted follow-up” in the message. We respond within five business days.

This is the planned intake for the marketing site. It is not a managed security operations center. Production deployment introduces an independent disclosure process.