Image: GitHub
Why it mattersA team already running static scanners can now feed those results into an AI that writes patches. The release is the first to turn external scanner imports on without a flag.
OpenAI's security tool for coding agents is getting bigger. On 10 November, OpenAI released version 0.9.1 of Codex Security. The tool is a CLI and TypeScript SDK that scans a repository for security problems. It then validates each finding and writes a patch. The npm package is @openai/codex-security. The repository on GitHub has 11,793 stars and 944 forks, under Apache-2.0. It was created in July 2026.
Version 0.9.1 is OpenAI's own release on GitHub. The release notes list three changes that matter.
What is in the release
Deep Scan now runs several Standard reviews in parallel batches. OpenAI writes that "workers can share leads and verify each other's evidence". The final report groups findings that would share one fix, while keeping the source location from each worker. The flag --stop-after-no now counts consecutive Standard reviews that add no findings, which is how you cap the budget.
External scanner imports are on by default in the CLI, the SDK, and the plugin. Before this release a team had to set CODEX_SECURITY_ENABLE_EXTERNAL_IMPORTS to use the feature. OpenAI writes that users can "preview supported reports, save findings locally, and publish supported findings to Cloud". The old environment variable is ignored.
The npm package now carries Developer ID-signed native add-ons for Apple Silicon and Intel Macs. Codex Desktop can load them. The scans logs command finds sessions faster in large compressed histories.
What the tool does
The README says Codex Security "helps you find, investigate, and fix vulnerabilities in your code". It runs four scan modes: Standard, Deep, targeted paths, and git-change-only. For each finding it decides whether the issue is real, writes a patch, and, OpenAI says, verifies an existing fix. It exports SARIF, JSON and CSV, writes a SECURITY.md, and opens tickets in Linear, Jira or GitHub. It supports OpenAI, Amazon Bedrock, OpenRouter, Fireworks AI and Cheaper Inference as the model provider.
OpenAI does not state a false-positive rate or an accuracy score in the README. Pricing is not set out as a table. One example in the GitHub Action documentation names an "estimated $5 limit per scan". That is OpenAI's own figure for the example, not a benchmark.
The cost of running it
Codex Security is a model-driven tool. Deep Scan asks the model to run several reviews, which means several sets of input tokens per repository. A team running this on every pull request should price Standard scans against their current scanner. Treat Deep Scan as a separate budget.
Source
openai/codex-security v0.9.1 on GitHub. Repository: openai/codex-security.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
