Lilith.
⌕
Editorial illustration: Codex Security puts scans, evidence and fixes into one workbench
Lilith illustration · editorial remix

OpenAI has demonstrated the Codex Security plugin in the desktop app and Codex CLI. A user selects a repository, revision, scope and model, runs a scan, then reviews the results in a dedicated security workbench.

One workflow now carries a finding from scan to verified fix

A standard scan leaves source code unchanged. It saves a readable report, individual findings and a coverage record. For each issue, the user is expected to inspect source evidence, attacker prerequisites, the validation method and remaining uncertainty.

The plugin can review a whole repository, selected paths or changes against a Git revision. Follow-up workflows can propose a patch and verify a fix. A separate package also provides a CLI, TypeScript SDK and CI integration. It requires Node.js 22.13.0 or a newer supported line and Python 3.10 or later.

Security review moves closer to the developer's daily loop

The practical shift is continuity. A developer does not have to move a finding between tools, restate the context and rebuild the evidence for review. The same session retains scope, report and remediation proposal.

Teams also get structured records. Outputs include a scan manifest, findings JSON and a coverage file. That supports archiving and automation instead of leaving a Markdown report as the only trace.

A completed scan still does not certify safe code

The documentation explicitly warns that a completed run can have coverage gaps and zero findings do not rule out vulnerabilities. Validation may rely on source analysis without executing a reproducer. A person who understands the system and its threat model still has to judge the result.

The demo offers no independent figures for accuracy, false positives or cost per scan. The recommended gpt-5.6-sol model at xhigh reasoning may help quality, but regular scanning economics cannot be inferred without operational data.

Repeatability and evidence quality will decide the product's value

Teams should measure whether the same commit produces stable findings, how many candidates security engineers reject and whether proposed fixes pass tests. Maintaining precise scope across multi-repository scans will matter as well.

Only a comparison with the current process can show whether the plugin shortens the path from finding to fix. The video shows a smooth interface. A queue of real pull requests will provide the harder verdict.

Lilith's verdict

Codex Security has seated the scanner beside the developer and handed it a folder of evidence. The real test is how many findings reach merge and how many land in the bin after the first human review.