Security is central to a tool that touches sensitive data. This page describes the controls the Service implements today. We make no claim to certifications we do not hold.
How your data is handled with AI
The core principle: the full dataset — all rows and all values — is never sent to any AI provider. The AI’s only job is to author rules (translate a schema hint or your plain-English instruction into parameters). Every result — every reconciliation break, data-quality score, match and profile statistic — is computed on our server by deterministic code. An AI cannot return a number it was never given.
- What actually reaches a model — only column names plus up to five sample values per column, used for mapping and suggestions. Before they are sent, a local PII scan redacts emails, SSNs, card numbers, IBANs, names, addresses, dates of birth and phone numbers.
- Never used for training — your content is never used to train AI models.
- Regulated-data mode — for BFSI / regulated data the Service can be set to route AI calls only to providers under a data-processing agreement with zero retention (AWS Bedrock, OpenAI or Anthropic), excluding free tiers that may retain prompts. It fails closed — AI features return a clear error rather than falling back to a non-DPA provider.
- Run it in your own environment — self-host the platform, or run Bedrock inside your own AWS tenancy, so even the masked schema samples stay within your cloud.
- Reference documents stay local — when you upload a data dictionary or rulebook for the Copilot to search, it is embedded and matched entirely on our server with a local model; that text does not leave the box.
Honest caveat: PII masking is pattern/heuristic-based — strong, but not a guarantee that an unusual custom identifier could never appear in a sample value. For the strictest requirements, use regulated-data mode or self-hosting.
Data isolation & access
- Per-tenant isolation — each user's workspace, datasets and remembered rules are separated; users see only their own data.
- Role-based access control — admin, analyst and read-only roles gate sensitive actions.
- Authenticated sessions with server-side session verification.
AI safety
- Bring your own key — you can route AI calls through your own provider account.
- Prompt-injection boundary — untrusted data (cell values, files, documents) is treated as data, never as instructions, so it cannot override the system's behaviour.
- Output filtering — model outputs are scanned and leaked secrets are redacted before display.
Integrity & monitoring
- Tamper-evident audit log — a hash-chained record of key actions that detects edits or deletions.
- AI health monitoring & drift evals — model calls are monitored and periodically evaluated for safety regressions.
- Dependency scanning in the build pipeline.
Transport & hosting
Traffic is encrypted in transit via TLS. The Service is hosted on Railway, which runs on Google Cloud Platform infrastructure; data at rest is protected by the platform's default disk encryption. For specific data-residency requirements, a dedicated hosting region can be arranged on the Enterprise plan.
Responsible disclosure
Found a vulnerability? Please email support@ai-datavalidator.com and allow us reasonable time to remediate before any public disclosure. Thank you.