Home / Solutions

Ship Codex changes through enforced reviews

Secure Codex AI coding agent use means setting sandbox limits, approvals, connections and enforced code reviews.

Updated November 16, 2026

Codex AI coding agent edits and runs code

OpenAI's Codex reads files, edits code and runs commands on laptops or in cloud environments. ZeroTB's software and team keep SOC 2 controls working after audits for US software, AI, fintech and healthtech companies.

  • Limit commands. The sandbox and approval policy decide what a task may run

  • Review connections. Apps, browser activity, web search and connected tools extend beyond local command limits

  • Review code. Codex Security examines application code for vulnerabilities that need investigation and repair

Codex security starts with sandbox and approval settings

The sandbox limits command access. OpenAI's Agent approvals and security explains the separate approval policy that governs requests for more permission.

Check permitted reach. Review which files and network connections the sandbox allows commands to access.

Review the patch. Permission to run a task leaves its resulting patch subject to company review.

Record the environment. Laptop and cloud permissions can differ. Record the environment and credentials Codex uses.

Connected services widen what Codex can reach

Connected services have access outside local command limits. OpenAI's Codex security cookbook covers apps, browser activity, web search and connected tools separately.

Check account permissions. The connected account determines which records and actions Codex can reach.

Review new connections. A new connection can expand access while the sandbox configuration stays unchanged.

Inspect active accounts. Check the account each connection uses, including the permissions granted to it.

Your data rules limit what Codex can read

The approved Codex task determines which company information it needs. Company data rules govern whether that information can be supplied.

Include useful context. Repository context may help Codex understand the source file it needs to change.

Justify sensitive access. Customer records and production credentials need a separate reason for inclusion.

Assess both sides. Vendor assessment covers data handling, while internal review determines what Codex may receive.

Codex Security finds vulnerabilities in the code itself

Codex Security examines application code for vulnerabilities. Its findings give the team issues to investigate and repair in that code.

Assign each finding. A finding gives its owner a specific issue to investigate in the code.

Understand accepted changes. The person accepting a patch needs to understand its changes and tests.

Keep change records. Record who reviewed the patch, which tests ran and how the release was approved.

Approved pull requests can hide missing branch rules

For SOC 2 change management, branch protection enforces required reviews or passing checks. Its coverage includes the receiving branch and bypass routes.

Test required review. An approved pull request can enter a branch that also accepts unreviewed changes.

Check new repositories. Agent pull requests can reach a new repository before its protections are set.

Recheck restored rules. A restored review rule can disappear again while the repair ticket still looks complete.

Release workflows decide which Codex changes reach production

A reviewed Codex change still needs an approved release route. Workflow permissions determine which people and service accounts can change production.

Follow approved routes. A reviewed change still needs to reach production through the agreed workflow.

Limit service accounts. Their permissions determine which production changes they can make.

Recheck workflow edits. Changes to workflows or service accounts can alter the route into production.

ZeroTB restores review rules after Codex changes

ZeroTB starts read-only, checks GitHub against SOC 2 commitments daily and prepares repairs. It checks systems Codex changes without connecting to coding agents.

Approve scoped repairs. Your team reviews and merges repairs. Write access is scoped by action and revocable.

Verify before closure. Fresh passing checks close gaps, while unknown, stale and disconnected results stay visible.

Decide acceptable exceptions. Your control owner decides whether a departure from the commitment is acceptable.

Change and development controls

ZeroTB covers the gaps Codex safeguards leave

The sandbox governs tasks, Codex Security examines code, and branch rules govern merges. ZeroTB takes responsibility for investigating and repairing control gaps.

TaskSandbox and approvalsCodex SecurityBranch rulesZeroTB
What it coversWhat tasks run and reachVulnerabilities in application codeBranch protection for accepted mergesSOC 2 commitments checked daily
What it leaves openAccess through connected servicesWhether required review is enforcedBypass routes and new repositoriesCode vulnerabilities remain Codex Security's job
Who acts on gapsWhoever configures the environmentThe team investigating each findingThe repository ownerZeroTB prepares repairs, your team approves
What auditors seeYour record of task scopeFindings to investigateThe current branch ruleA signed Verified Control Record

Example of blocking unreviewed merges again

Here, tests confirm that a restored branch rule blocks unreviewed merges. The Verified Control Record keeps the requirement, condition, change, approval and result.

Keep unreviewed changes out
A repository review record: an unreviewed change merged despite the policy, the owner's approved correction, unreviewed changes blocked and reviewed ones merging, and the evidence kept with the approval.
  1. Missing review rule. The repository accepts an unreviewed change despite a policy requiring another reviewer.

  2. Approved rule change. The owner approves restoring required review on the branch receiving changes.

  3. Verified merge conditions. An unreviewed change is blocked, while a change with the required review proceeds.

  4. Retained closure evidence. Observed states and verification results remain beside the approved correction.

Check five things before Codex reaches production

Define approved work

Establish where Codex runs, its permitted tasks and the data it can reach.

Inspect access boundaries

Check files, services and the accounts behind each connection.

Check change enforcement

Inspect repository and release rules for required reviews and tests.

Approve and verify repairs

Obtain the owner's approval, then test the corrected condition.

Keep evidence current

Keep the repair's decision and verification together, with an owner for further checks.

Sources

  1. OpenAI, Agent approvals and security, accessed November 4, 2026.
  2. OpenAI, Codex Security, accessed November 4, 2026.
  3. OpenAI, Codex security cookbook, accessed November 4, 2026.

Start with the Pilot

Discuss Codex controls. Over eight weeks, the Pilot maps up to 25 commitments and takes two real gaps through verified restoration.