Built the way careful people build things.
Curiosity comes out of years running artificial intelligence in places with no tolerance for sloppy data handling, where a mistake had real consequences. The same posture sits underneath everything on this site, whoever you are and wherever you are signing in from.
What we put in place by default.
Six controls applied at the database, the application, and the edges of every integration. None of them are optional.
Account isolation
Postgres row level security policies make sure every read and every write is scoped to the account that asked for it. Reaching another account's data is impossible at the database itself, not just in the application above it.
Agents do not delete
The artificial intelligence agent has no authority to delete records or files anywhere in the system, across the web chat, every add in, and every surface. Deletion is a human action with a confirmation step. Enforced in code, audited at runtime.
Outlook safety
The Outlook add in reads and proposes drafts. It cannot send, forward, delete, move, or modify any email. Email bodies have URLs stripped and HTML entities decoded before any model sees them, so prompt injection attempts from external senders are kept out of the model.
Token budgets on every call
Every call to a model writes a usage row into the observability pipeline. Monthly token budgets keep runaway costs from buggy loops or abuse contained. The per request limits are deliberately generous so legitimate work is never cut off.
Prompts are versioned and audited
Every change to a system prompt is versioned, attributable to a named user, and auditable. A bad prompt can be rolled back in two clicks. The safety rails, including the Outlook posture, the deletion policy, and attachment handling, live in code rather than inside any editable prompt.
Secrets never reach the browser
The API keys for AWS Bedrock, Azure OpenAI, Stripe, and Voyage live on the server only. The client side bundle is built to never carry them. Source maps are stripped from production builds.
If something goes wrong.
A standard incident response playbook adapted for an artificial intelligence workspace. Notification timelines are firm commitments, not aspirations.
- 01Detect the issue, through monitoring or a customer report
- 02Triage and assess impact within the first hour
- 03Contain the issue and remediate the cause
- 04Notify any affected customer within twenty four hours
- 05Write a post incident review and send it to the affected customer inside five business days
Where the compliance work sits today.
Australian Privacy Principles and the Essential Eight are already in place. The certified frameworks are in the queue and dated against real targets.
Enterprise customers can request a SIG Lite questionnaire response and a security summary covering hosting, sub processors, and the incident response policy.