Codex Security: install the plugin, run your first scan, review every change
A step-by-step Codex Security walkthrough captured from a real screen recording: install the plugin in the ChatGPT desktop app, run a first scan with gpt-5.6-sol at xhigh effort, step up to deep scans, and put a security review in front of every merge.
Codex Security in five lines
- Codex Security is OpenAI's application security agent: it reviews code for vulnerabilities and reports evidence you can use to investigate and fix them. It ships as a ChatGPT plugin for local repos, plus a CLI, a TypeScript SDK and a research-preview cloud.
- Install it from the Plugins page in the ChatGPT desktop app (or /plugins in Codex CLI), then start a new chat on the repository — the quickstart's rule is to never continue an existing chat.
- For the best scan quality the docs specify gpt-5.6-sol with xhigh reasoning effort. The first scan is read-only: your source stays unchanged while the agent validates plausible findings.
- Deep scans run multiple independent standard scans over the same repo and combine their findings — reduced variability, much longer runtime. For diffs and pull requests, use $codex-security:security-diff-scan instead.
- Every scan leaves report.md, findings/<slug>/, hardening/ and JSON manifests behind. Keep the directory intact, and check the coverage summary before you trust a clean result.
Codex Security: Make Your App Safer
Channel: Software Engineer Meets AI3:59
Codex Security — OpenAI Codex documentation
Docs: developers.openai.com
Frames come from a full screen recording of the Codex Security plugin in use — no facecam and no burned-in captions on any kept frame. Facts are cross-checked against the official Codex Security documentation (plugin quickstart, deep scans, code changes) and the openai/codex 0.163.0 and 0.164.0 release notes. The recording shows model chips like 5.6 Sol, matching the docs' current gpt-5.6-sol recommendation.
Video frames remain the property of their creators and are embedded here as step-by-step documentation with attribution and deep links.
From install to your first Codex Security report
Meet the security agent
- 1
Know what Codex Security does before you install it
Codex Security is OpenAI's application security agent: it reviews code for vulnerabilities and reports evidence you can use to investigate and fix them. It ships in four shapes — a plugin for local repositories in Codex, a CLI for repeatable terminal scans and CI, a TypeScript SDK, and a research-preview cloud that watches connected GitHub repositories. Everything in this guide uses the plugin.
The docs frame Codex Security as an agent for finding, confirming and fixing vulnerabilities across plugin, CLI, SDK and cloud surfaces. - 2
Understand the Analyze, Understand, Threat Model loop
Every scan runs the same loop. Codex analyzes the repository, builds an understanding of how data and trust boundaries actually flow through your code, then generates a threat model that decides what deserves scrutiny. This is why a scan takes minutes rather than seconds — and why the docs tell you to scan only code you own or have permission to assess.
The video's hand-drawn model of a scan: Analyze first, Understand second, Threat Model last. - 3
Install the Codex Security plugin from the Plugins page
In the ChatGPT desktop app, open Codex, then open Plugins and search for Codex Security. The public plugin is listed as Security scanning for your codebase — install it, confirm it is enabled, and a Security entry appears in the sidebar. On the CLI the same move is /plugins, search, Install plugin, then /new.
Typing codex se into Plugins search surfaces the public Codex Security plugin card.
Run your first scan
- 4
Start a fresh chat on the repository you want assessed
The quickstart's rule: open the repository in Codex and start a new chat — don't continue an existing one, or stale context pollutes the scan. Pick the model the docs call for: gpt-5.6-sol with xhigh reasoning effort. The scan is read-only; it leaves your source unchanged and saves a report with evidence and remediation guidance.
The plugin quickstart covers install, a first read-only scan, and reviewing the result in Codex. - 5
Set up the scan: Codebase or Changes, scope, context
Starting a scan opens a Set up security scan dialog. Scan type is Codebase for a repository or folder, or Changes to review a PR, commit, or patch. Verify the scan area (entire codebase or one folder), the current branch and last commit, and optionally add threat model scoping guidance so the agent prioritizes the surfaces you care about. Then press Start scan.
The scan dialog locks the target to codex-security-demo-commerce on branch main before anything runs. - 6
Read the findings: severity, coverage, evidence
A completed scan opens a Findings workspace with severity counts, coverage (28/28 files in the demo) and one row per reportable finding — here a High-severity authorization bypass / IDOR on an order detail endpoint. Each finding carries source evidence, attacker prerequisites and validation, and validation can use source analysis without running a reproducer.
The demo repository's first scan surfaces one High IDOR finding across 28 files scanned.
Go deeper with a deep scan
- 7
Decide between a standard scan and a deep scan
A standard scan is one review pass over the repository or a scoped folder — right for first runs and routine checks. A deep scan runs multiple independent standard scans over the same target and combines their findings, trading a longer runtime and more tokens for reduced variability between runs. Deep scans cover repositories and folders only; diffs stay on the change-review workflow.
The docs' comparison table: deep scans trade runtime and resources for reduced variability. - 8
Kick off a deep scan from the chat composer
The conversational way in: send Use $codex-security:deep-security-scan to run a deep security scan of this repository. For one component in a monorepo, replace this repository with the path to its folder. Codex opens the scan setup first and waits for your Start scan action before it inspects anything.
The deep-security-scan prompt typed into a fresh Codex chat on the clicksis repository. - 9
Confirm setup, then let preflight run
The Deep Security Scan skill opens the setup workspace: Deep scan enabled, scan area set to Entire codebase, current branch and last commit verified, optional threat model scoping guidance. Only your Start scan action releases the plugin to run its preflight against the runtime before the discovery phase begins.
Deep scan setup waits for an explicit Start scan before the repository is touched. - 10
Watch the phases: preflight, discovery, review, report
Deep scans run repeated discovery plus centralized validation and reporting phases without modifying repository files. The Scan progress panel shows where you are — preflight checks, mapping the attack surface, reviewing code, validating findings, tracing impact, building report — and how many parallel scans are active.
Step 3 of 4: preflight finished, discovery running, six scans active in the background.
Review changes and keep the evidence
- 11
Review uncommitted changes and PR ranges
Before you merge, scan the diff, not the repo. For uncommitted work: Use $codex-security:security-diff-scan to review my current uncommitted changes for security regressions. For a branch range, name both revisions and add a focus list: Focus on authentication, authorization, input handling, filesystem access, network requests, and secrets.
Two copy-paste prompts cover working-tree changes and origin/main-to-HEAD ranges. - 12
Keep the scan directory: report.md and the JSON artifacts
Every completed scan writes report.md as the readable entry point, a findings/<slug>/ directory with one detailed vulnerability report per finding, hardening/ proposals when the scan has reportable findings, and structured scan-manifest.json, findings.json and coverage.json for automation. Keep the whole directory together when sharing so the links from report.md keep working.
report.md is the entry point; findings, hardening and the JSON manifests sit beside it.
Daybreak: the trust tier behind serious Codex Security work
The 0.163.0 release of Codex CLI (October 15, 2026) turned Daybreak on by default for eligible accounts — and if you run security workflows through Codex, that matters. Daybreak is OpenAI's Trusted Access for Cyber program: an approval-based access tier that relaxes model refusals for verified, authorized security work.
- 1Daybreak Blue is the defensive tier: reduced refusals on mainline models — GPT-5.5, GPT-5.6 Sol, GPT-6 Sol, GPT-6 Luna — for verified defensive work such as vulnerability triage, secure code review, detection engineering, incident response and patch validation. It's the recommended starting point for most security teams.
- 2Daybreak Red is the offensive tier: access to cyber-specialized models like GPT-5.5-Cyber plus reduced refusals on Astra, for authorized penetration testing, red teaming and exploit validation. It needs separate approval and is limited to business and enterprise organizations.
- 3In Codex CLI you enable it with /daybreak in interactive sessions or daybreak=true in headless commands. The 0.163.0 release notes state it is enabled by default for eligible accounts, and codex exec can select it with an OpenAI API key when the server offers access.
- 4Approval alone doesn't flip anything — Daybreak must also be enabled for the request through your workspace or API project controls, and the toggle sits off by default in the Codex app's advanced model picker.
- 5Scope is strict: authorized defensive work on systems you own or are explicitly permitted to test. No customer-facing applications, no third-party access, no resale of access — standard usage policies still apply on top.
- 60.164.0 (October 23, 2026) kept those defaults and only polished the edges — Daybreak notices now render after the session header — so the practical takeaway is to stay current on CLI releases and treat Daybreak as the tier you request when Codex Security triage keeps stalling on legitimate findings.
For most readers the default posture is already right: Codex Security runs fine on standard safeguards with gpt-5.6-sol at xhigh effort. Daybreak exists so approved teams don't stall mid-triage when a legitimate exploit path needs full analysis — request it when your workflow, not your curiosity, needs it.