Skip to content

Security

Every action has a boundary.

AI can plan the work. Audoryn decides whether an external action is allowed, must wait for a person, or must stop.

Deterministic authority, human oversight

Three possible answers. None of them decided by the model.

Choose a decision to see what happens to the same request.

audoryn / action gatewayIllustrativeREC-3117

Proposed by the model

Create an issue

Worker
Developer assistant
Target
GitHub · product repository
  1. Agent identityThe worker and its organization are confirmed
  2. Declared capabilityThe operation matches a granted scope
  3. Risk and current policyCurrent conditions and the deterministic rule
Held for review

The action waits for a person.

The run pauses and an authorized reviewer receives the request. Before an approved action resumes, Audoryn checks its authority again.

The record

  1. 09:12Research analystRead issues in the product repositoryAllowed
  2. 09:31Operations coordinatorPost an update to #leadershipHeld for review
  3. 09:47Developer assistantDelete a branchDenied

The control path

A connected tool does not grant permission to use it.

Each request passes explicit checks before it reaches an external system. AI instructions and memory give context; they never become authority.

  1. Who

    Identity

    Confirm the worker and its organization.

  2. What

    Capability

    Match the requested operation to a declared scope.

  3. Now

    Risk and policy

    Evaluate current conditions against the deterministic rule.

  4. Outcome

    Act, wait or stop

    Allow, require human approval, or deny.

  • Allowed
  • Held for review
  • Denied

Operational safeguards

Control continues after a decision.

Authority can change while work is underway. Audoryn keeps the relevant checks and the human response in the execution path.

  • 01

    Distinct identities

    A person's session cannot act as an AI worker. Every managed worker has its own agent identity, and organization boundaries are checked.

  • 02

    Protected credentials

    Connected provider credentials stay inside the backend credential boundary. Models never receive provider tokens or direct provider tools.

  • 03

    Human intervention

    Sensitive actions can wait for review. An organization, worker or integration can be suspended without deleting its configuration.

  • 04

    A record to inspect

    Requests, decisions, approvals and provider outcomes carry correlation context, so operators can follow what happened.

The record

What happened, who acted and why.

Every decision leaves an entry that connects the request, the rule that applied and the result.

TimeReferenceActorActionPolicyDecisionOutcome
09:12REC-3115Research analystRead issues in the product repositoryRead access to approved repositoriesAllowedCompleted · 14 issues read
09:31REC-3116Operations coordinatorPost an update to #leadershipLeadership channels need a reviewerHeld for reviewApproved by Daniel K. · posted
09:47REC-3117Developer assistantDelete a branchDestructive repository actions are not grantedDeniedBlocked · provider not called

IllustrativeIllustrative records · example data

Security by design

Useful autonomy needs clear authority.

See how these controls fit into the product, then start with a worker and a defined job.