Announcing Control Tower

Cross-posted from LessWrong.

An open-source framework for running and comparing AI control research

We built Control Tower, an open-source framework built on Inspect for running AI control evaluations, and the Control Leaderboard, which compares control protocols and monitors across environments.

Coding agents make it faster for researchers to build environments for their own research, so they can test new control techniques quickly. However, each new paper runs on its own environments, monitors and scoring, so control techniques are hard to compare or replicate across papers. Also, once a paper is published, its code is rarely maintained, so its techniques seldom get tested again on new environments or models. How can researchers keep moving quickly with their AI control research agendas, while enabling others to replicate and build on their results?

With Control Tower, our team ports researchers' environments and techniques into shared evaluation infrastructure, so research results can be replicated, compared and built on. This post outlines how we (1) port research, (2) compare metrics and (3) update the framework with new models and features, as well as our roadmap.

(1) We port settings and protocols from existing research

Most AI control experiments are similar. A setting gives the agent an environment to operate in, with an honest job to do (main task) and a secret attack to attempt (side task). A protocol is the blue team's defense, such as a monitor that flags suspicious actions (see docs for more info). Given the similarities, we can port settings and protocols from existing research into the Control Tower framework.

Control Tower began as the code for LinuxArena, and was built on UK AISI's ControlArena. As LinuxArena grew, we replaced ControlArena with simpler evaluation code. In March 2026 we split that code out as Control Tower, so other settings could run on it too. In September 2026 we added protocols that act while the agent runs.

As of November 2026, Control Tower has ten settings and four protocols. We started with LinuxArena. Four settings came from ControlArena: SHADE-Arena, vLLM code sabotage, SAE interpretation sabotage and the APPS backdooring setting from AI Control. We ported five more from the research: BashArena, Refactor Arena, ResearchArena, container sandbox breakout (still being validated) and PostTrainBench (private). Three of the protocols are defer to trusted and trusted editing, from AI Control, as well as defer to resample from Ctrl-Z. We also run Watcher, from Apollo Research, in two ways: as a protocol that blocks actions live, and as a monitor that scores each action. As we port each new setting, we add the required features to Control Tower.

(2) Compare models, protocols and monitors plus download the data

The Control Leaderboard compares models, protocols and monitors by safety and usefulness. Safety is the share of attacks caught, out of all attacks, at a fixed audit budget (e.g., 1%). Usefulness is the share of main tasks a model completes when it isn't attacking. The leaderboard has three types of evals: models, protocols and monitors.

Model evals show how useful each frontier model is, and how often its attacks get past a trusted monitor. For example, GPT 5.4 Nano completes 49.0% of its main tasks but only 24.6% of its side tasks, which makes it a weak attacker (79.3% safety at a 1% audit budget).

Control evals compare protocols, such as trusted monitoring, by how much safety they add and how much usefulness they cost. For example, with DeepSeek V4 Flash as the agent and GPT 5.4 Nano as the monitor, trusted monitoring raises safety from 0% with no protocol to 72.8%, and costs no usefulness.

Monitor evals tell you which monitor setup best separates honest actions from attacks, by running every monitor on the same transcripts. AUROC measures how well a monitor ranks attacks above honest runs, from 0.5 (chance) to 1.0 (perfect). For example, on BashArena, GPT 5.4 Nano at high reasoning effort (0.59 AUROC) nearly matches Claude Haiku 4.5 with no reasoning (0.62), at under a third of the cost per task.

Most trajectories can be downloaded as Inspect .eval files from the Data & downloads page.

(3) We maintain the framework and the leaderboard, so ported work can be rerun as new models and environments arrive

We maintain Control Tower and the leaderboard, so ported settings and protocols keep running after their papers are published. Each major release makes Control Tower easier to install, run and upgrade. For example, v6 installs from PyPI, pulls every setting with one command, runs samples up to 11 hours on EC2, and runs new monitors on existing transcripts (v6 release notes).

When new models come out, we add them to the leaderboard. When a setting changes, we rerun only the samples it affects, so results stay comparable.

Roadmap

Next, we're focused on:

  • Supporting multi-agent control evals where the untrusted system is several agents, such as an agent that delegates to subagents, agents running at the same time, or one attack spread across several tasks. Each action will record which agent took it, so monitors can watch each subagent.
  • Updating the Control Leaderboard with more data including more monitors and protocols.
  • Adding and replicating more control papers including replications in ct-replications and adding new papers along with required features (backlog).
  • Facilitating future research including Auto Control, a setting where AI agents develop Control Tower itself, and improving onboarding and documentation.

Try it out and share your feedback

Try out Control Tower and let us know how we can make it more intuitive and useful. You can request we add research, request a feature or report a bug.

Thanks Adam Hanson, Tomáš Turlík and Radek Budzyński for feedback on this post.