Every request is checked against the customer, role, task, and approved use before the assistant receives context.
JTF Copilot docs
Scoped AI for security workflows.
Build agents and artifacts that respect client scope, authorized methods, human approval, and audit requirements.
Docs index
The control surface is part of the product.
Requests for another client, unauthorized records, covert surveillance, spyware, interception, trackers, or false identities are refused and audited.
Generated reports, proposals, and summaries are saved with the requester, reviewer, approval status, and source notes.
The API accepts authorized AI requests, returns approval decisions, and uses the same review rules as the workspace.
Guardrail examples
R-AI-1 and R-AI-2 are hard refusals.
Draft an after-action summary from records the requester is authorized to see.
Requests for another client's records are refused.
Requests for covert surveillance, spyware, or another client's records are refused.
Requests include clientIds, allowedMethods, prompt, requester identity, and optional human approval state.
Refused requests return scope_violation, out_of_bounds_method, or human_approval_required before artifact creation.
Prompt content is reduced to a hash and the decision is persisted for review without exposing unrelated client context.