Approach
AI Agent Security Architecture
The Custos concept organises our engineering approach into five areas of control. It is an architecture discussion, not an assertion of a generally available hosted product.
Five areas to review
Content and context
Retrieved documents, messages and tool output may carry instructions that conflict with the user’s task.
Action control
Define permitted actions, resource boundaries and approvals outside the model.
Agent-native surfaces
The agent’s dependencies and stored context form part of its security boundary.
Human oversight
Route consequential actions to someone who can understand and authorise the proposed change.
Assurance and evidence
Connect requirements, test results and operational records without unnecessarily copying sensitive data.
From design to delivery
For each engagement, establish which controls already exist, which need engineering, the compatible tooling and the acceptance tests. Product availability, hosting, support coverage and deployment commitments must be confirmed in the proposal.
See a bounded demonstration
A small executable example shows allow/block decisions for tool arguments. It is a local policy demonstration, not a deployed security platform or evidence of production readiness.
Start with a clear scope
Tell us which systems, actions and review requirements are in scope. We will discuss the work, responsibilities and deliverables before you commit.