Least privilege
A user or agent reaches only the functions and data its task requires, according to its role.
Trust is built into the product
This center explains the principles that guide how we design our systems. It is not a claim of any security certification or government accreditation unless one is stated explicitly with verifiable evidence.
Separated product and customer data
Multi-tenant systems are designed so one organization’s data cannot move to another.
Traceability and review
Sensitive operations and important automated decisions are logged so they can be reviewed and investigated.
People make the decision
Sensitive legal or administrative outputs are never approved without the authorized person.
Grounded in sources
When an answer can be documented, it is linked to its source, and quoted text is kept apart from conclusions.
Clear about limits
We explain the nature of simulation and AI and never present them as an official authority or absolute certainty.
Responsible AI
How do we limit hallucinations and risk?
Sources before wording: Retrieving documented data takes priority over generating an answer.
Evidence checks: Links between a claim and its source are reviewed, and uncertainty is shown when evidence is missing.
Execution limits: Consequential actions need clear permission, and human approval can be required before execution.
Continuous measurement: Quality, cost, and operational errors are monitored, and behavior is reviewed when the model or data changes.
Report a security or privacy issue
Do not send passwords, access keys, or sensitive personal data in your first message. Name the product and give a short description of the problem.