Participants build a monitor that flags suspicious actions taken by a coding agent, test ways an attacker might evade it, and measure its performance across many examples.
Suggested time: 90 minutes
Exercises: Open the participant instructions
| Area | Prerequisites coming in | Main learnings going out |
|---|---|---|
| Engineering | - | Run ControlArena evaluations, inspect trajectories, and turn monitor outputs into analyzable scores |
| ML | Confusion matrices, precision, recall, and decision thresholds | Interpret score distributions, ROC curves, AUC, thresholds, and capability-gap effects |
| Security | Detector errors and adaptive attackers | Write a monitor policy, design an evasion policy, and evaluate both on held-out behavior |
| Theory | Trusted models, untrusted models, and monitors in AI control | Explain what monitoring assumes about trusted models, observability, and attacker knowledge |
- Read BlueDot's Introduction to AI Control.
- References: the ControlArena documentation and scikit-learn's ROC/AUC guide.