Action authorization
Should this agent action proceed?
Evaluate the actor, execution context, requested target, approval state, and deterministic policy before a governed side effect proceeds.
Make AI actions and data release deliberate before they happen.
Agent Authority puts a governed decision boundary in front of agent actions and data release. It combines authenticated identity, bounded authority, deterministic policy, and decision-linked evidence.
Zero Trust authority
Authorize the action. Authorize the release.
No agent, runtime, or network position receives standing authority. Each governed action is evaluated against authenticated identity, request context, deterministic policy, and required evidence. An action may be permitted while its data must be narrowed, approved, or blocked for a particular destination.
Semantic Action Review
Does this action require accountable human review?
A specially governed assessor returns bounded risk evidence about a proposed action’s meaning and context. Deterministic policy decides whether that evidence requires review; the assessor cannot authorize, deny, execute, or modify the action.
Data-release authorization
Should data leave this boundary?
Evaluate data with PII, secret, and customer-provided classifier signals in the context of the intended release.
Decision evidence
Can the outcome be explained later?
Keep the request, policy basis, semantic assessment, decision, and resulting outcome connected for review and investigation.
Governed path
Decide before the side effect, not after the incident.
1
Enter through a governed path
Hosted workspaces and supported external integrations present actions at an enforcement boundary instead of treating a log as the control.
2
Bind authority to this request
Evaluate an authenticated participant and specific route rather than relying on broad ambient runtime access.
3
Apply authoritative policy
Deterministic policy can allow, deny, narrow, or route a request for approval. Required enforced checks fail closed when their result cannot be obtained.
4
Record the decision and outcome
Maintain attributable evidence so teams can investigate both blocked and allowed activity.
Agent DLP
Protect data at governed egress boundaries.
Built-in PII and secret checks, plus customer-provided filters or classifiers, provide evidence for policy decisions at governed model, tool, channel, service, and external-agent boundaries.
Separate concern
Prompt injection is not a DLP synonym
Inbound model-input defenses are distinct from PII and secret egress controls; the appropriate coverage depends on the governed boundary.
Policy remains authoritative
Signals inform the decision
Detectors and classifiers provide evidence. They do not independently publish policy or override deterministic denies.
Hosted and external paths
Use the right boundary for the runtime you operate.
Hosted execution
Built-in governed runtime path
Atellagent hosts the workspace execution surface, giving it the most direct enforcement and observation boundary.
Supported external hosts
Coverage follows exposed host hooks
The public client can govern documented Codex, Claude Code, and Gemini CLI hook points. It cannot claim control over lifecycle events that a host does not expose.
Review the boundary in the context of your environment.
Read the architecture for the technical model or schedule a review of your action and data-release paths.