Claude Code vs Codex: sandbox, approvals and what neither tests

Codex sandboxes local work by default, with the network off, and asks before leaving the sandbox. Claude Code starts in auto mode, where a classifier reviews actions, and its OS sandbox is opt-in. Both ship branch reviews and deep repository scans. Neither hosts your app or tests it once deployed, so database policies, keys in the bundle and per-user access stay yours to check.

The team behind Keybrum's attacksUpdated 11 min read

In short

  • Codex's default for a version-controlled folder is a sandbox that writes only in the workspace, with network access off. Claude Code's sandbox is off until you run /sandbox.
  • Claude Code's default guard is auto mode, a classifier that blocks actions it judges unsafe. Codex has a similar reviewer, Auto-review, that you turn on yourself.
  • Both vendors publish their flaws. OpenAI issued three Codex CVEs on 1 September 2026, fixed in CLI 0.131.0 and the matching desktop builds; Anthropic lists more than 30 advisories for Claude Code.
  • Codex Security and Claude Security scan source code. OpenAI's own FAQ says a scan with no findings "can still miss vulnerabilities".
  • Neither tool sees a table left open in a dashboard, a secret behind a NEXT_PUBLIC_ prefix on your host, or account B reading account A's data. Test the deployed app.
On this page

The short answer

Claude Code is Anthropic's coding agent. Codex is OpenAI's: a CLI, an IDE extension, the ChatGPT desktop, web and mobile apps, and cloud tasks, included in every ChatGPT plan from Free to Enterprise according to OpenAI's pricing page. Both read your repository, run commands and write code.

Their security models rest on the same two layers: a sandbox that limits what a command can touch, and an approval policy that decides when the agent must ask. The defaults differ. Codex leans on the sandbox first. Claude Code leans on a classifier first. This page compares the two layers, the review tools and the cloud setups, from each vendor's documentation as read on 13 November 2026. It does not rank them on coding quality: we have no data on that, and it is not what leaks your users' data.

For each tool alone, see Claude Code security and Codex security.

Claude Code vs Codex: security features side by side

QuestionClaude CodeCodex
Who makes itAnthropicOpenAI
Default on your machineAuto mode from version 2.1.283: a classifier reviews actions. OS sandbox offAuto preset in a version-controlled folder: workspace-write sandbox with on-request approvals. Read-only elsewhere
OS sandboxOpt-in with /sandbox. Seatbelt on macOS, bubblewrap on Linux and WSL2. Not on native WindowsOn by default. Seatbelt on macOS, bwrap with seccomp on Linux and WSL2, a native sandbox on Windows
Network from commandsWith the sandbox on: a local proxy with an allowlist that starts emptyOff by default in the CLI, the desktop app and the IDE extension
Protected pathsWith the sandbox on: .claude settings and shell startup files stay write-denied.git, .agents and .codex are read-only even inside the workspace
Automatic reviewer of risky actionsAuto mode classifier, on by defaultAuto-review, set with approvals_reviewer = "auto_review". The default reviewer is you
No-guard modebypassPermissions--dangerously-bypass-approvals-and-sandbox, also --yolo
Rules the model may missCLAUDE.mdAGENTS.md, 32 KiB by default
Rules that can blockDeny rules and PreToolUse hooks.rules files (experimental) and PreToolUse hooks
Branch review/security-review/review, and the Codex Security plugin for a security scan
Pull request reviewCode Review, Team and EnterpriseCode review on GitHub, and Security Review on Pro, Business, Edu and Enterprise
Cloud runsOne isolated VM per session, GitHub credential kept outside itOne VM per task, network secrets swapped in by a proxy
Hosting or database for your appNoneNone
Sources: code.claude.com/docs and OpenAI's Codex documentation at learn.chatgpt.com/docs, read on 13 November 2026.

Sandbox and approvals: two different defaults

Codex: the sandbox comes first

Codex has three sandbox modes. read-only lets it inspect files. workspace-write lets it read, edit and run commands in the working directory. danger-full-access removes the filesystem and network limits. The approval policy sits on top: on-request asks when Codex needs to leave the sandbox, never does not ask but still stays inside the sandbox mode.

In a folder under version control, Codex starts in the Auto preset: workspace-write with on-request. In a folder that is not, it starts read-only. Network access is off by default, and commands the agent starts, such as git or a package manager, inherit the same limits. OpenAI warns that turning the network or web search on exposes the agent to prompt injection, and tells you to treat web results as untrusted.

Claude Code: the classifier comes first

Claude Code starts interactive terminal and VS Code sessions in auto mode. A separate classifier model reviews actions and blocks ones such as curl | bash, sending sensitive data to outside endpoints, production deploys, force pushes and mass deletions. Local file operations in the working directory and installs from your lock file are allowed. Anthropic's docs say auto mode "reduces permission prompts but does not guarantee safety".

The OS sandbox is a second layer, and you turn it on with /sandbox. Then shell commands write only to the working directory, a temp directory and folders you added. Reads stay open on most of the machine, including ~/.ssh, until you add deny entries. The sandbox covers shell commands only: file tools, hooks and local MCP servers run outside it. On native Windows, commands run unsandboxed.

Review and scan tools: what each one reads

ToolVendorWhat it readsAvailability
/security-reviewAnthropicThe changes on your current branch, one passBuilt into Claude Code
Security-guidance pluginAnthropicWhat Claude writes: each edit, each turn, each commit it makesAll plans
Claude Security pluginAnthropicA repository or a diff, with verified findings and patches you applyPaid plan, API access or a cloud provider
Code ReviewAnthropicPull requestsTeam and Enterprise
/reviewOpenAIA branch diff, uncommitted changes or one commit, without changing your filesCodex app, CLI and IDE extension
Codex Security pluginOpenAIA local repository, read-only, with a report and remediation guidanceInstalled from Plugins
Security ReviewOpenAIA GitHub pull request's diff and supporting code. Also on demand with @codex security reviewPro, Business, Edu and Enterprise. Not Plus
Daybreak Defense FactoryOpenAIA connected GitHub repository, new commits, scheduled scansPreview, billed by token usage

Both vendors describe the limits themselves. OpenAI's FAQ for its cloud scans says that "a scan with no findings can still miss vulnerabilities", and its plugin page says a finding can be validated by source analysis without running a reproducer. Anthropic says scans are nondeterministic, so two scans of the same code can surface different findings, and that its review reads the source code in your checkout, "not a running site or deployed service".

One practical difference: Codex Security Review posts findings on the pull request, where anyone who can view the pull request can read them. OpenAI's docs point this out for public repositories. Think about that before you enable it on an open-source project.

Secrets and cloud runs

On your machine, neither tool hides your .env file by default. In Claude Code you add deny rules such as Read(./.env); its auto mode allows reading .env and sending a credential to its matching API. In Codex, the sandbox limits writes and the network, and OpenAI's sandbox page does not describe a default block on reading files. With both, the working rule is to keep production keys off the machine the agent runs on.

In the cloud, the designs differ in a useful way. Each Claude Code cloud session runs in an isolated VM managed by Anthropic, and your GitHub credential never enters it: a proxy attaches it on Anthropic's side. Each Codex cloud task also runs in a VM. Codex separates environment variables, passed directly to programs, from network secrets: the program gets a placeholder, and a proxy swaps in the real value only for the HTTPS destinations you allowed. If a cloud task needs your Stripe or Supabase key, the network secret is the safer of the two.

What neither of them checks: the deployed app

Neither Claude Code nor Codex hosts your app, runs your database or signs your users in. Both write files, and everything they set up lives in your accounts. So the leaks that reach production are the same with either one, and most are outside what a code scan can see.

  • Database policies. A table created by a migration without row level security, or a policy loosened by hand in the Supabase dashboard. Firebase rules edited in the console. No repository file changes, so no scan fires.
  • Keys in the bundle. A secret stored on Vercel or Netlify under a NEXT_PUBLIC_ or VITE_ name is published in your JavaScript. The repository only holds the variable's name.
  • Per-user access. A route that checks the session and forgets the owner. A scan can flag the pattern; only two real accounts prove whether B reads A's invoice.
  • What is live. An old branch, a preview pointed at production data, source maps, a served .env file, a Stripe webhook that accepts unsigned events.

The first item has a number behind it. CVE-2025-48757 was assigned after 170 of 1,645 Lovable apps tested had missing or weak row level security. The cause is a Supabase table without a policy, and a migration written by Claude Code or Codex can produce the same table. See the Supabase RLS guide for the policies and vibe coding security risks for the wider pattern.

A self-check for an app built with either

Four checks, on your own project only.

  1. 1

    Review the branch with the agent you use

    In Claude Code, run /security-review. In Codex, run /review against your base branch, or ask for a Codex Security scan. Fix what is real. This is the cheap part.

  2. 2

    Ask the database, not the code

    Run the query below in the Supabase SQL editor. It lists every policy, the role it applies to and its condition. A qual of true on a table with user data means every caller with that role reads every row.

  3. 3

    Look at what the browser receives

    Open your live site, then the network tab, and search the JavaScript files for sk_, service_role and PRIVATE KEY. The frontend API key guide lists which keys are public by design.

  4. 4

    Replay requests as someone else

    Run the commands below against your own app: once with a second account's cookie, once with none. The IDOR guide explains what each answer means.

sql
select tablename, policyname, roles, cmd, qual
from pg_policies
where schemaname = 'public'
order by tablename, cmd;
bash
# Your own app only. As account B, ask for a record that account A created: expect 404.
curl -i https://your-app.example/api/invoices/ID_CREATED_BY_A \
  -H "cookie: SESSION_COOKIE_OF_B"

# Signed out: expect 401.
curl -i https://your-app.example/api/invoices/ID_CREATED_BY_A

# Stripe webhook with no signature: expect 400.
curl -i -X POST https://your-app.example/api/webhooks/stripe \
  -H "content-type: application/json" -d '{}'

Which settings to change first

  • Claude Code: run /sandbox and deny reads of .env, ~/.ssh and ~/.aws/credentials.
  • Claude Code: keep the workspace trust dialog, and remember that claude -p shows no trust prompt.
  • Codex: stay on the Auto preset. Use --sandbox workspace-write for scripts, as OpenAI recommends, and leave --yolo for disposable machines.
  • Codex: leave network access off unless a task needs it, and prefer network secrets over environment variables in cloud environments.
  • Both: put your security rules in CLAUDE.md or AGENTS.md, then back the ones that must hold with a PreToolUse hook. OpenAI calls its hooks "a useful guardrail, not a complete enforcement boundary".
  • Both: update. Codex CLI should be at 0.131.0 or later, Claude Code on its current version.

Where Keybrum fits, and where it does not

Keybrum tests neither agent. It has no view of your terminal, your sandbox mode or your approval policy, and it does not replace /security-review or a Codex Security scan. Those read code before a merge. Keybrum attacks the app after the deploy, and only when its owner has proved ownership.

It works from the outside, the way a stranger would. It lists the Supabase tables your public key can reach and reads them with no account, then as a new account. It collects every API route your JavaScript calls and calls each one with no session. It creates a marked record as one account and requests it as another. It checks whether keys found in the bundle or in API answers work, sends one unsigned event to the Stripe webhook, posts a short prompt to AI routes with no session, and requests /.env and source maps. It also proves Next.js flaws such as the middleware bypass by behaviour. A finding always carries the request that worked.

Claude Code and Codex both speak MCP, so either can read the findings and the fixes from Keybrum's MCP server (setup), apply them and ask for a retest. A one-time report costs €49 for one app. Plans start at €29 a month excluding VAT for 5 apps, and attack again on every deploy.

Questions

Is Codex safer than Claude Code?

Out of the box, Codex restricts more on your machine: a sandbox with the network off in any version-controlled folder. Claude Code's default is a classifier, and its sandbox is opt-in. Once you turn Claude Code's sandbox on, the two are close. Both have published sandbox or approval bypasses, so neither default is a promise.

Does Codex have a command like /security-review?

Codex has /review for a general review of a branch, uncommitted changes or a commit. For security specifically, OpenAI ships Codex Security: a plugin that scans a local repository, a CLI for CI, and Security Review on GitHub pull requests, which you can request with @codex security review.

Does either tool test my live app?

No. Both review and scan source code. Anthropic says its review reads your checkout, not a running site or deployed service, and OpenAI's Codex Security pages describe scans of repositories, commits and pull requests. Policies set in a dashboard and behaviour across two accounts need a test of the running app.

Which plans include Codex?

OpenAI's pricing page says Codex is included in ChatGPT Free, Go, Plus, Pro, Business, Edu and Enterprise. It can also run in the CLI, SDK or IDE extension with an API key, billed at API prices and without cloud features such as GitHub code review.

Can a malicious repository attack me through these agents?

It has happened with both. Codex's CVE-2026-19592 let a prepared repository run a program through Git's core.fsmonitor setting without a prompt, and Claude Code's CVE-2026-33068 let a committed settings file skip the trust dialog. Both are fixed. Keep the tool updated and read a repository's agent configuration before opening it.

Do AGENTS.md and CLAUDE.md enforce security rules?

No. Both are instructions the model reads, and a model can miss an instruction on a long task. Anthropic says CLAUDE.md is context, and points to hooks when an action must be blocked. Use the file for rules and a hook or a CI check for the ones that must never break.

Sources

  1. 1.Agent approvals and security, OpenAI
  2. 2.Sandboxing, OpenAI
  3. 3.Codex Security, OpenAI
  4. 4.Codex Security Review, OpenAI
  5. 5.Daybreak Defense Factory FAQ, OpenAI
  6. 6.Code review, OpenAI
  7. 7.Cloud environments, OpenAI
  8. 8.Hooks, OpenAI
  9. 9.Pricing, OpenAI
  10. 10.CVE-2025-59532, CVE Program
  11. 11.CVE-2026-19591, CVE Program
  12. 12.CVE-2026-19592, CVE Program
  13. 13.Security, Anthropic
  14. 14.Choose a permission mode, Anthropic
  15. 15.Sandboxing, Anthropic
  16. 16.Catch security issues as Claude writes code, Anthropic
  17. 17.Scan your codebase for vulnerabilities, Anthropic
  18. 18.Security advisories for anthropics/claude-code, GitHub
  19. 19.CVE-2026-33068, CVE Program
  20. 20.CVE-2025-48757, NVD

Your app is public. So are its mistakes.

Check what any attacker can already see in it, free, in ten seconds. The full attack, with proof and fixes: €49 once, or €29 a month on a plan.

or continue with Google or GitHub