Control Tower Documentation

Control Tower is an AI control evaluation framework for measuring the effectiveness of safety protocols when deploying potentially misaligned AI agents. It provides realistic software environments where AI agents complete tasks while safety protocols monitor for malicious behavior, and models AI control as a game between a Blue Team (maximizing usefulness and safety) and a Red Team (attempting to subvert safety measures).

Where to start

• Running evaluations on existing settings: start with the Quickstart, then the User Guide covers running evaluations, monitoring, viewing trajectories, and measuring safety and usefulness.
• Building settings and tasks: start with the Hands-On Guide, then the User Guide covers environments, porting a benchmark, main tasks, side tasks, scorers, sabotage trajectories, and protocols.

Both paths use the objects defined in Core Concepts; the Glossary spells out every abbreviation. To change Control Tower itself, see CONTRIBUTING.md.

Stability

From v6.0.0, only a major release makes a breaking change, and a breaking change is a change to the stable API:

• the ct run, ct runs, ct settings, ct protocols, ct live and ct login commands, as the CLI reference lists them;
• these Python modules, and the registry entries the Python reference lists:
  • policies and protocols: control_tower.eval.red, control_tower.eval.blue, control_tower.eval.white, control_tower.eval.harness, control_tower.eval.game_view and control_tower.eval.registry;
  • monitors: control_tower.monitors and control_tower.monitors.registry;
  • tasks and samples: control_tower.task.task, control_tower.task.space and control_tower.task.sample_metadata;
  • results: control_tower.trajectories.sample, control_tower.metrics.audit and control_tower.metrics.control, plus only suspicion_scores and SuspicionScore from control_tower.trajectories.monitor_score and only Cost and Priced from control_tower.models.pricing;
• the names Your Own Task lists for building your own Inspect Task, imported from the modules it gives;
• the eval log format;
• the env API setting packages build on (ct-setting-sdk and its wire protocol), versioned on its own: only an SDK major changes it incompatibly.

A listed module's stable names are the public names it defines, or its __all__ when it has one, and not names it only imports; the types those names take and return are stable with them. Everything else under control_tower is internal and can change or move in a minor release. ct extras and the modules behind it, control_tower.extras and control_tower.datasets.importers, are not part of the stable API either and can change or move in any release.

User Guide

Basics

Concepts

Running

Protocols and Monitors

Analysis

Building Environments

Tasks and Scorers

Sabotage Trajectories

Reference

Command Line, Python, and HTTP

Data and Glossary

Project Structure