Skip to content

Review skills

A skill is an instruction pack that guides the agent’s review. The tutorial uses security-review-repo. The alternatives below let you try another approach. Compare their purpose and limits in the tool shortlist.

Create the VM and complete the tutorial’s environment checks before installing a pack. Run installation commands on your host machine, then ask the agent to use the installed skills inside the VM. Read third-party packs before installing them. To repeat a run with the same instructions, select a fixed Git commit rather than a branch that may change. The installation reference covers file selection, installation paths, updates and troubleshooting.

security-review-repo adapts Anthropic’s MIT-licensed security-review prompt to review a whole repository instead of pending changes. Its source is in sandbox/skills/security-review-repo.

Run this command on the host to install the skill directly from GitHub:

Terminal window
appsec-sbx skills appsec-sbx https://github.com/aai-institute/agentic-appsec-playbook --subdir sandbox/skills

The --subdir sandbox/skills option selects the directory containing security-review-repo. The wrapper downloads the repository to a temporary checkout on the host and removes that checkout after preparing the files for transfer. The VM does not need access to GitHub.

By default, the command installs from the main branch. To install a specific revision, add --ref <commit> with the Git commit you want to use. The wrapper records the installed revision in skills.json on the host.

To install a version you have edited locally, or a pack from a private repository, give the command the path to the pack in your local checkout:

Terminal window
appsec-sbx skills appsec-sbx /path/to/agentic-appsec-playbook/sandbox/skills

In OpenCode inside the VM, ask: “Use the security-review-repo skill to review this repository.” The report is written to ~/out/findings.md. Follow the tutorial’s review and export steps.

Google’s Mantis provides skills for several stages of a security review. Its npx skills add installer needs access to GitHub, which the VM’s review network policy blocks. Clone the repository on the host, select a commit you have reviewed, and copy the pack into the VM with appsec-sbx skills:

Terminal window
git clone https://github.com/google/mantis /path/to/mantis
git -C /path/to/mantis checkout --detach <reviewed-commit>
appsec-sbx skills appsec-sbx /path/to/mantis

Replace <reviewed-commit> with the Git commit you selected.

Mantis’s reproduction and patching stages expect Docker inside the agent’s environment. The workload user in this VM cannot use Docker, so limit the run to the stages that work with source text and record which stages you skipped.

Mantis writes working files into the imported repository. Before repeating the review or comparing it with another tool, export the results, stop the VM and revoke the old credential at the provider. Then reset the VM, import the target and reinstall the skills. See Starting another review.

Run this command on the host to install the skills from the repository’s main branch. The --subdir option selects its .claude/skills directory:

Terminal window
appsec-sbx skills appsec-sbx https://github.com/anthropics/defending-code-reference-harness --subdir .claude/skills

For repeatable runs, add --ref <commit> with the revision you have reviewed. Check that the selected revision provides the commands and report names below.

Inside the VM, open Claude Code and enter this command to start the scan:

/vuln-scan /home/appsec/target/source

The /vuln-scan skill reviews the source and writes VULN-FINDINGS.json and VULN-FINDINGS.md in the target directory. After the scan, exit Claude Code. Run the following command in the VM shell to copy both reports into ~/out, the directory included by appsec-sbx export:

Terminal window
cp ~/target/source/VULN-FINDINGS.json ~/target/source/VULN-FINDINGS.md ~/out/

The reference harness’s autonomous pipeline and Docker environment require separate setup. The appsec-sbx skills command installs only the skill pack. For the discovery exercise, finish by exporting the raw reports. Triage, reproduction and patching belong to later stages of the review.