Purpose
Himaya uses AI models to understand intent in messages, classify sensitive data and explain its decisions. This policy sets out how customer data may and may not be used with those models.
No training on your content
Your content is never used to train shared or general-purpose models, by Himaya or by the model providers we use. Models process your data only to deliver the service to you.
Your email, files and findings never train a model used by another customer.
Learning that stays in your tenant
Some features improve as you use them. That learning belongs to your tenant alone:
- Communication baselineThe relationship graph of who your people talk to is built for your tenant and used only for it.
- Your feedbackConfirmed false positives adjust future scoring for that sender within your tenant.
- Trainable classifiersExample documents you upload to teach Himaya your sensitive content are used only for your tenant, and only at your request.
Model providers
Where we use third-party models, we use them under terms that do not permit the provider to train on customer content. Providers that process customer data are listed as subprocessors in our Trust Center.
Human control
AI informs decisions inside limits you set:
- Deterministic firstDeterministic checks run before models. Hard evidence, such as a malicious sandbox result, takes precedence over a model's opinion.
- Policy decidesData loss prevention actions are made by your policy, with no model in the decision step.
- Your boundariesAutomated remediation runs only within the limits your team defines. Everything else waits for approval.
- ExplainedEvery verdict records the evidence and reasoning behind it, and can be reviewed and reversed.
Model safety
Text from inspected messages, attachments and web pages is passed to models as data, never as instructions. When a model is unavailable or uncertain, Himaya does not mark content as safe. It falls back to review.
Review
This policy is reviewed at least annually and whenever we add a model provider or change how models use customer data.
