Security, sovereignty, and accreditation.
Where Syntasa runs, what it depends on, and what has been accredited. Zero trust is only enforceable when you can see your AI — this is where you see ours.
Deployment models
Syntasa deploys as a single codebase across four substrates. Software provenance is identical in all four; only the promotion path differs.
Air-gapped operation is the design baseline. The other three models are relaxations of it — not the reverse.
No route to the internet at any time. All dependencies vendored. No callbacks, no telemetry, no external model APIs. Google Distributed Cloud certified; deployed in the nation's highest classified environments.
Zero-trust principles
Identity
Every request — human or agent — is authenticated against the customer's identity provider. There are no shared service accounts and no Syntasa-held credentials in production.
Privilege boundaries
Least privilege is enforced at the data, model, and agent layers independently. An agent's authority is a declared, reviewable artifact — it cannot exceed the identity that invoked it.
Auditable data movement
Every read, transformation, and export is logged with actor, time, and justification. Audit records are append-only and exportable to the customer's SIEM.
Ingest-time governance
Governed at ingest: lineage, classification handling, and policy enforced from the moment data arrives, not after.
Contained by design. Observable by default.
Agent containment
Autonomous agents change the failure model. When an agent holds credentials and a route out, the agent's reach — not your policy — becomes the effective boundary of your data. The industry's recent agent incidents share one shape: broad authority, standing credentials, and no perimeter between the agent and the outside world.
An agent inside the perimeter cannot leak what the perimeter does not release.
Syntasa agents run inside the same sealed boundary as the data, under the same controls as any human operator. Containment is a property of the architecture, not an instruction the agent is trusted to follow.
AI observability
Containment answers what an agent can reach. Observability answers what it actually did. Syntasa gives full visibility into what models and agents do with data — every model, pipeline, job, query, and service on one governed plane, with the causal chain behind each output preserved.
Accreditations, with scope
Jurisdiction and data handling
Customer data is processed only within the customer's chosen jurisdiction and boundary. Syntasa operates deployments in the US, UK, and EU. Data residency is structural: the platform has no mechanism to move customer data to Syntasa infrastructure, because no such infrastructure sits in the data path.
Syntasa personnel access production systems only under customer-granted, time-boxed, logged credentials. Support does not require data access; where diagnostic data is shared, it is customer-initiated and customer-reviewed.
One sentence version: your data never touches our systems.
Disclosure policy
Security researchers may report suspected vulnerabilities to security@syntasa.com. We acknowledge reports within two business days, and we do not pursue action against good-faith research conducted within scope.
Customers under active agreement receive notice of confirmed, in-scope security incidents per contractual terms — typically within 24 hours of confirmation.
Need this in writing?
Accreditation letters, architecture documentation, and security questionnaires are available under NDA.