xorcise-skills

Agent skills for XORCISE — the platform for benchmarking AI agents on hands-on cybersecurity missions.

Each skill under skills/ is a self-contained SKILL.md plus its scripts, usable by any AI coding agent that supports the skill format. More skills will be added over time.

Skills

Requirements

These skills drive XORCISE — they do not replace it, and they will not install it for you.

• A running XORCISE engine — xorcise-ai/xorcise, with its REST API reachable (default http://127.0.0.1:3001). See docs.xorcise.ai.
• Docker, with permission to build images and run containers.
• /dev/net/tun, so a sandbox can join the mission network.
• python3 + PyYAML.
• An API key for each model provider you want to benchmark.

Per-skill prerequisites live in the skill's own README.md.

Installing a skill

Place the skill's directory where your agent discovers skills, then invoke it by name. For agents that load skills from ~/.claude/skills/:

cp -r skills/xorcise-playbooks ~/.claude/skills/
# …or symlink it, to keep tracking this repo
ln -s "$PWD/skills/xorcise-playbooks" ~/.claude/skills/xorcise-playbooks

The skill is then available as /xorcise-playbooks.

See the skill's own SKILL.md and README.md for the full walkthrough and requirements.

What leaves your machine

• Your API keys stay on your host. They are typed into a chmod 600, git-ignored file and read from it directly. Provider variables are handed to the sandbox by name, so docker inherits the values and the skill never reads them.
• One key is visible to local process listing. OpenHands takes a single LLM_API_KEY, and the runner passes that one value on the docker run command line — so it can be read from /proc/<pid>/cmdline by other accounts on the same machine while a run is in progress. Use a dedicated host. See SECURITY.md.
• Prompts and agent activity go to the model provider you configure — and only to that provider. Their terms govern that data.
• Telemetry goes to your own XORCISE instance, over OTLP (default port 4318) on the host running the benchmark.
• Reports stay local. Eval cards are written wherever you point --outdir and are never uploaded. Keep them outside a git checkout — run artefacts contain agent transcripts and mission flags.
• XORCISE.AI receives nothing. No vendor telemetry, no licence check, no phone-home.

Full detail in SECURITY.md. Our Privacy Policy covers the website and hosted platform, not these skills — they run entirely on your machine.

Security

Never report a security vulnerability in a public issue. Use GitHub's private advisory form at https://github.com/xorcise-ai/xorcise-skills/security/advisories/new, or email [email protected]. Scope, the disclosure key and what to include are in SECURITY.md.

These skills run untrusted agent code against deliberately vulnerable targets. Run them on infrastructure you are willing to lose — see the operational note.

Contributing

Issues and pull requests are welcome. Start with CONTRIBUTING.md — it covers how a skill is laid out, how to test a change, and what gets a pull request rejected.

Participation is governed by the Code of Conduct. Maintainers are listed in MAINTAINERS.md.

This GitHub repository is canonical. Development up to the first public release happened in a private repository, which is now archived.

Licence and acceptable use

Licensed under the Apache License 2.0, © 2026 Fifth Domain Pty Ltd (ACN 606 251 585), trading as XORCISE.AI. NOTICE carries the attribution and trademark terms. Contributors sign the CLA — see CONTRIBUTING.md.

XORCISE is offensive-security tooling. Only point it at systems you own, or that you have specific written authorisation to test. The Acceptable Use Policy sets out what that means in practice, including export-control and sanctions obligations. That policy is not an additional licence condition and does not narrow the rights Apache-2.0 grants you.

Use of the XORCISE name and logo is governed by the Trademark Policy.