Privacy at the model layer isn’t a feature you can add later. It’s a posture you decide on day one. Here’s how we approached it.
The threat model
Customer data flowing through an AI workload can leak in three places: the request, the model’s training data, and the response logs. Each has its own controls, and “we’re SOC 2” is not a sufficient answer for any of them.
What we actually do
Customer prompts and responses are encrypted in transit and at rest. They never enter our training pipeline. They’re scoped to the customer’s tenant, with audit logs available on request.
The hardest part wasn’t the technology. It was building the discipline to say no when product asked for “just a little” cross-tenant analytics.
Compliance + something more
SOC 2 Type II, GDPR, HIPAA, ISO 27001 — we’re certified across the board. But the more useful question we ask ourselves: would we let our own engineers read this data? If the answer is no, our customers shouldn’t have to either.
/ JOIN THE CONVERSATION
Leave a comment.