Security and data handling

QI has not been assessed against SOC 2, ISO 27001, FedRAMP, CMMC, NIST SP 800-171 or any export-control framework, and we make no claim of compliance with any of them. If a customer needs one of those, QI is not ready for that customer yet.

• Classified information.
• ITAR- or EAR-controlled technical data, and controlled unclassified information (CUI), in any environment operated by QI.
• Anything you are not authorised to share. A pilot uses customer-authorised, non-controlled, historical campaign records.

The team is based in Hyderabad, India. Where a customer requires data to stay in their own boundary, the intended model is a deployment inside the customer's environment rather than QI-operated hosting. That deployment model is a design goal, not a finished, certified offering.

• Authenticated access with roles (admin, engineer, reviewer, viewer). Records are isolated per organisation on the server, and cross-organisation access is covered by automated tests.
• Sign-in sessions are tracked on the server, expire after a set time and can be revoked, so signing out ends the session rather than only hiding it. Repeated failed sign-ins are throttled.
• Original uploaded files are stored unchanged and hashed (SHA-256). A file is never overwritten.
• The audit trail is append-only: the database rejects edits, deletions and truncation of audit events for the application. A database administrator with owner rights could still change it, which is why a deployment should run the application under a restricted database role.
• Artifacts can be given a retention period. A job the operator schedules removes expired originals, extracted text and exports built from them and records the purge. Backups are outside that job.
• Uploads are validated by content, not by file extension. Macro-enabled and embedded-object Office files, mismatched file types, oversized files and suspicious archives are rejected.
• Documents are parsed in a separate process that has no database credentials or API keys, with a time limit.
• Every extracted fact keeps the exact location it came from, and the original file is re-read before an evidence package is exported.
• Nothing is exported while proposed links or findings still await a named engineer's decision.

• AI assistance is off by default and has to be switched on by whoever operates the deployment.
• When it is on, suggestions must quote the source text verbatim and are checked against it. Anything unsupported is discarded. Accepted suggestions stay pending until an engineer reviews them.
• AI never approves a requirement, closes a nonconformance, or declares anything qualified.
• QI does not train models on customer data.
• The AI path has been exercised against scripted test providers only. It has not been measured against a real model, and we do not claim real-model accuracy.

• No single sign-on or multi-factor authentication yet, and no per-campaign membership: access is by organisation and role.
• The browser keeps the sign-in token in session storage and no content security policy is configured yet.
• Request limits are shared across several servers only when the database-backed limiter is configured; they are not a substitute for a network-level control.
• Encryption at rest and key management depend on how a deployment is configured; the prototype does not ship its own key management.
• No independent penetration test has been performed.
• The public demo on this website uses synthetic data only and is not connected to any customer system.

A dedicated security mailbox has not been set up yet. Until it is, please use the contact form on the home page and mark the message as a security report. Please do not include sensitive data in that message.