Trust Center ยท Architecture
Six layers, each assuming the one outside it has failed.
Defence in depth applied to agent systems, where the risk extends beyond caller identity to what an agent is permitted to do once authorised.
Edge and ingress
Request filtering, rate limiting and schema validation before anything reaches an agent. Untrusted input never arrives at a model call unscreened.
Identity and entitlement
Every caller, whether human or agent, carries a scoped identity. Entitlement is resolved per request against the domain the Brain is permitted to read.
Network micro-segmentation
Service mesh with mutual TLS between workloads. Lateral movement between services is denied by default rather than filtered after the fact.
Model safety and guardrails
Semantic guardrails and PII screening on both the prompt and the completion, with policy applied before a tool call is allowed to execute.
Data and cryptographic storage
Encryption at rest and in transit, tenant-scoped keys, and tokenised handling of sensitive fields so raw identifiers are not stored in the graph.
Agent sandbox runtime
Agents execute inside bounded runtimes with explicit tool allow-lists, execution budgets and refractory limits that stop runaway loops.