Agent Guard CI

Agent Guard CI is my attempt to watch for security problems with the same principles as my code architecture tests. I don't want agents to be able to edit the CI they are being tested with. The scanners live in their own repository and install into my cluster as one shared Argo Workflows template, so the application code being checked never contains the checks. If an agent runs into a problem with a check, my fleet's rule is that it banks a work item and a decision for me instead of changing its CI.

A run clones the repository, runs the scanners in parallel, and fails if any finding scores 7.0 or higher on the CVSS severity scale. Agent Guard CI has nine components:

• TruffleHog finds secrets committed by mistake, such as API keys and tokens.
• Semgrep finds insecure patterns in my own code.
• KICS finds misconfigurations in infrastructure code.
• Socket scores how risky each dependency is, to catch supply-chain attacks such as malicious look-alike packages.
• Syft and Grype list every package the code uses and check that list against known vulnerabilities.
• Pulumi CrossGuard checks my infrastructure code against policy rules.
• Two optional uploaders send findings to DefectDojo for triage and keep the raw reports in storage.
• A policy gate reads every report and decides pass or fail. If its threshold is missing or isn't a number, it refuses to run instead of passing everything. It fails closed.

The same checks run as pre-commit hooks and again in CI, because hooks are local and easy to skip. I also wrote down what Agent Guard CI does not cover yet: image signing, provenance, admission control, runtime detection and testing the running app. Knowing what I am missing and need to improve is just as important as what is built.

I built the first version in about a week in April 2026 and refined it through May. The workflows are Argo Workflows and are tied into my Argo Workflows migration. So they will go live once a repo moves off of Forgejo.