Security and deployment
Moflet AI isolates each tenant, keeps secrets out of the model and keeps an audit trail you can sign. The same product runs in the cloud, in a dedicated perimeter or air-gapped.
Authorization at every boundary
The PDP is consulted at the gateway, before invoking an agent, before every tool and before reading a credential. It is least privilege at every hop, including the graph, not a screen-level role. Every row carries tenant_id and the database enforces RLS. That is two isolation layers per tenant.
Secrets never reach the model
The vault encrypts with a per-tenant KEK and treats the secret as write-only. It is injected at the tool transport. It does not appear in the prompt, the log or the trace. When data must be forgotten, the design includes crypto-shredding.
Every decision leaves evidence
Audit is append-only: no update and no delete. Every event has an actor. The export can be signed and, optionally, hash-chained. The semantic layer adds a minute per query, tied to the run. What security asks for in the review is in the record, not in a CSV afterwards.
Three perimeters, one product
Multi-tenant: logical isolation, many organizations on one platform. Dedicated: the same chart, a single tenant, when the contract needs its own perimeter. Air-gap: no egress, with internal SMTP, self-hosted models and identity on your network. The API, the PDP and the audit model do not change between the three.