LinuxArena

Welcome to Control Tower, a framework for AI control evaluations built on Inspect and developed by EquiStamp, Redwood Research, and Luthien.

🚨 Control Tower is in active development. Star this repo to follow updates, or Watch > Releases to get notified on new versions.

Found a bug or have a question? Open an issue, submit a PR or book a call

Control Tower provides many built-in components, including blue-team protocols, action- and trajectory-level monitors, control evaluations, safety-usefulness curves, and trajectory tooling. Environments and their main and side tasks come from separate settings repositories, and extensions (new protocols, policies, monitors, and scorers) can be provided by other Python packages.

See the documentation for an overview of Control Tower, a quickstart, and development guides.

To get started:

git clone https://github.com/linuxarena/control-tower && cd control-tower
uv sync
alias ct='uv run ct'        # add to your shell profile to keep it
cp .env.example .env          # set ANTHROPIC_API_KEY (or OPENAI_API_KEY)
uv run python scripts/generate_model_rates.py   # model prices, so evals can show what they cost
ct settings pull linuxarena

Check the install without spending a model call. It takes a couple of minutes:

ct run eval --policy ct/test --task-file simple --tag test

Then run a real eval. The first one builds the environment image, which takes about ten minutes:

ct run eval --policy ct/attack -e linuxarena/clinical_trial -t adverse_reaction_followup_tracking -s suppress_safety_notifications

---

See CONTRIBUTING.md for code style, typing, testing, and pull request conventions, and the development setup guide for local services, environment pulls, and shell completions.