Security
Reduce the blast radius. Keep actions legible.
Atarla does not make absolute security claims. The useful promise is narrower: limit what any one component can reach, make actions reconstructable, and recover cleanly.
Design goals.
Least privilege, workload isolation, identity for every device and service, encrypted transport, brokered secrets, signed updates, explicit permissions, auditability, defined offline behavior, and responsible vulnerability reporting.
Proximity is not trust. A service, device, or model earns no implicit trust for being local or customer owned, which follows the zero-trust position that trust is not granted by network location or asset ownership.
Treating dependencies as dependencies.
External models, packages, agents, tools, connectors, and retrieved documents are dependencies, not trusted authorities. Untrusted content can influence what an agent proposes. It should not be able to widen what an agent is permitted to do.
Published guidance for language-model and agentic applications treats prompt injection, sensitive information disclosure, supply chain, improper output handling, and excessive agency as distinct risk classes. The white paper maps each to a location in the architecture.
What is not claimed.
Atarla is not certified and not audited by a third party, and no system is free of risk. This page describes design goals and architecture, not verified security outcomes.