Install and use the INTRUSIVE.AI skills
The public skills repository gives your AI harness instructions for preparing environments, operating campaigns, and reporting results. Installing skills does not install the products, configure credentials, or start a campaign.
Install for your assistant
Use Bash 3.2 or newer and Git on macOS or Linux. The installer can also run in WSL; this does not establish Windows support for the sandbox products themselves.
git clone https://github.com/intrusiveai/skills.git
cd skills
./install.sh --help
./install.sh --target codex --dry-run
./install.sh --target codex
Choose --target claude for Claude Code or --target both for both assistants.
These are alternatives, not additional required steps.
| Assistant | User installation | Project installation |
|---|---|---|
| Codex | ~/.agents/skills/ |
<project>/.agents/skills/ |
| Claude Code | ~/.claude/skills/ |
<project>/.claude/skills/ |
To install for one project, add --project-dir /absolute/path/to/project.
For another harness, use --dest /its/documented/skills/directory and enable
discovery through that harness. The installer copies files; it does not configure
an arbitrary assistant's discovery mechanism.
Open a new assistant session in your working project and confirm skill discovery. Keep the complete installed folders, including references and assets.
Choose a skill
| Installed name | Use it to |
|---|---|
interceptor-sandbox-management |
Install Interceptor, inspect an agent, prepare blueprints, and manage sessions and evidence. |
operator-sandbox-management |
Install Operator and compatible Attack Harness, prepare campaigns, execute, and inspect results. |
scenario-bundle-generation |
Structure your chosen objectives and scenarios using a target capability export. |
attack-simulation-results-reporting |
Create a portable HTML review from retained campaign evidence. |
campaign-debugging |
Investigate operational failures using native component evidence. |
All five install by default. To select a subset, use repeated --skill options
instead of the full installation command:
./install.sh --target codex --skill interceptor-sandbox-management --skill operator-sandbox-management
Repository folder names use underscores; installed names use hyphens.
Ask for a concrete outcome
Provide your workspace, target source, releases, provider access method, objectives, and limits when you have them. For example:
Use interceptor-sandbox-management to prepare this application in Interceptor.
Inspect dependencies, build its image, create and validate the blueprint, and
verify a legitimate application operation. Report remaining gaps.
Use scenario-bundle-generation to create a bundle from ./objectives.md
using ./operator-environment/capabilities.json. Preserve my objectives and do not
add new scenarios or a testing agenda.
Use attack-simulation-results-reporting to create a portable HTML report from
./exports/support-review. Account for untested scenarios and cite retained evidence.
Claude Code also supports the installed names as slash commands. Without native
skill support, make the folders available and ask the assistant to read the relevant
SKILL.md and references. Reading instructions does not grant terminal access,
credentials, or execution permissions.
These management skills belong to your assistant. They are not automatically instructions supplied to Attack Harness inside a campaign.
Update or remove
Installed skills are independent copies. Updating the checkout does not update them. Obtain the desired revision, move selected installed folders outside the discovery directories, then rerun the installer. Compare local edits first. The installer refuses existing selected folders; avoid duplicate versions in discovery.
To uninstall, remove only the selected installed folders and open a new session. Record the skills revision with setup notes when reproducibility matters.