Codex Security's GitHub Action opens scans to Bedrock, OpenRouter and Fireworks
Action 0.5.0 adds three inference routes and GitHub Enterprise Cloud support, with explicit model credentials and the existing Linux sandbox requirements.
OpenAI’s Codex Security GitHub Action can now run its security scans through Amazon Bedrock, OpenRouter or Fireworks. The October 20 Action 0.5.0 release adds those provider inputs alongside the default OpenAI route, giving teams another way to connect repository analysis to model access they already manage.
This is a change to the scanning workflow, not a new model. Each non-OpenAI provider requires an explicit model identifier and its own credentials. Bedrock uses AWS credentials and a region, with a documented GitHub OIDC role example; it does not require an OpenAI API key or Codex login. OpenRouter and Fireworks use their respective API keys. Model access and usage costs remain the user’s responsibility.
A supported route for Enterprise Cloud
The release also supports GitHub Enterprise Cloud repositories on *.ghe.com, including HTTPS and SSH checkouts. GitHub Enterprise Server remains unsupported. The documented setup uses a GitHub-hosted Ubuntu 24.04 x64 runner and requires Bubblewrap plus the relevant AppArmor configuration. Sandbox preflight must succeed before a scan runs.
Existing users can update their workflow to the release’s full Action commit, b46717e825c91fe01a350f496c3f6e4b6acfa457, after reviewing the setup. The distribution installs @openai/codex-security@0.5.0. New workflows should follow the complete versioned example, including permissions, sandbox preparation and credentials, rather than copying only the provider input.
Findings remain report-only by default; a severity threshold can make findings fail the job. Scanner and required reporting errors still fail it. The guide keeps provider credentials out of installation and dry runs, recommends checkout without persisted credentials, and limits its PR example to trusted same-repository contributions. A skipped fork PR has not been scanned.
The time settings still use different units
Deep Scan in the Action continues to use max-time-hours. The CLI’s new --max-time option uses minutes and is not an Action input. The discovery allowance also does not promise that final reporting finishes within that allowance.
The release notes additionally describe clearer artifact-upload errors and publish a distribution manifest and software bill of materials. These sources establish the shipped integration and documented behavior. AITrending has not run comparative vulnerability-detection tests or verified that switching providers preserves finding quality.
