How the pieces are meant to fit
Architecture
Three security directions across the life of an AI system: what it learns from, what it does while running, and what security events it produces.
Conceptual target architecture, not a statement of working cross-product integration.
Before deployment · data and templates
Implemented (Training Data Security)AI Security Audit
Examines what a model learns from: training data (poisoned samples, ranked review queue — implemented) and model templates (static analysis — in development). One product; Template Security is part of it.
While running · agent context, tools, actions, memory
Implemented (text screening)Runtime Security
Screens the text an agent reads, channel by channel, and asks a human to confirm risky outbound actions (limited). Memory integrity, MCP sessions, and tool-call policy enforcement are in development.
Across both · security events
In developmentSecurity Intelligence Layer
Ingests and normalizes security events, correlates them, tracks provenance, and is meant to surface cross-product insight. A local release candidate exists; production deployment and end-to-end integration are not proven.
The layers are drawn stacked, not connected: there are no working data flows between products to draw today. Where each capability stands is published with status labels on the Technical Evidence page.