05 — Secure gated agent preview access

This appendix is a vendor-neutral control pattern for a temporary review surface. It is not a deployment runbook, an account of current preview exposure, or evidence that a particular network boundary has been measured. A preview can reveal code, tool responses, files, logs, or task context; its access grant therefore needs independent review.

Control objectives

Objective Required decision and evidence
Named scope Record purpose, owner, approved reviewer identities, exact service and ports, review window, expiration, and teardown owner before granting access.
Separate identity Use a preview identity distinct from higher-trust operator and control-plane identities. Do not turn a preview path into a management path.
Least privilege Restrict entry at the access gate and the runtime's local network controls to approved reviewers and the intended service. Deny unrelated services and routes.
No network expansion Do not enable subnet routing, exit-node or relay behavior, public sharing, or remote shell merely to make a preview reachable. Any exception needs separate authorization and threat review.
Secret custody Keep join credentials and application secrets outside the repo, prompts, transcripts, logs, and receipts; reference protected secret storage rather than copying values.
Reversibility Define the steps and owner for closing the service, withdrawing the grant, revoking the temporary identity/credential, and removing temporary network state before opening access.
Evidence Record sanitized authorization, permitted and denied reachability checks, readiness, expiration, and inactive-state checks. A configuration file alone does not prove these behaviors.

A private overlay, access proxy, bastion, or comparable gate can satisfy the objectives when its actual policy and runtime behavior are verified. No particular vendor, daemon, interface, port range, firewall syntax, or topology is required by this appendix. Firewall changes must be designed against the existing ruleset and management path; a new input-hook chain with a global policy drop can interrupt unrelated traffic and is not a safe generic example.

Safe validation checklist

Perform this review in the approved environment with an operator-specific, separately reviewed procedure. Do not treat this page as authorization to change live access.

Stop and rollback criteria

Stop before opening access if ownership, approval, data classification, a bounded scope, the independent management path, or a reversible change plan is missing. Do not improvise a broad firewall rule or copy credentials into command lines to make the preview work.

Close the preview immediately and follow the approved rollback procedure if an unapproved identity or port can reach it, the service becomes publicly reachable, the management path is affected, a credential appears in output or a public artifact, the service differs from the reviewed artifact, or the expiration/teardown owner is lost. If rollback cannot be verified, keep the preview closed, preserve a sanitized incident record, and escalate to the operator; do not mark the session complete.

Publication boundary

A published example or receipt must use fictional values and omit live routes, identities, credentials, raw transcripts, and private topology. Run the repository's public safety scan and manually apply the declassification checklist before publication. A scanner result does not prove safe exposure or successful teardown.

This gate complements workload containment, default-deny egress, explicit promotion, audit, and separately demonstrated rollback. It does not establish production readiness, current isolation, or a completed recovery drill.